Best Local LLM Tools

Run, serve, and chat with models on your own hardware — the tools that actually deliver. Ranked by our usability scores across setup ease, docs, community health, and maturity — 174 tools reviewed in full.

01

OpenSpec

Fission-AI/OpenSpec · ★ 69k ▲518/7d · TypeScript

OpenSpec is a command-line workflow for writing requirements, design notes, and implementation tasks before an AI coding assistant changes a repository. It keeps those artifacts as Markdown in the project, so the plan can be reviewed, revised, and archived with the code instead of disappearing into chat history.

SetupDocsCommunityMaturity

Verdict Our OpenSpec run passed all 4,211 tests after a 16-second install, making it an unusually easy spec workflow to trial with an existing coding agent. Use it if vague chat instructions are already causing rework and your team will genuinely review the generated Markdown.

Full review →

02

llmfit

AlexsJones/llmfit · ★ 37k ▲1.3k/7d · Rust

llmfit is a terminal tool that detects your RAM, CPU, GPU, and local model runtimes, then ranks language models by whether they should fit and how they may perform. It helps you narrow a crowded model catalog before downloading large weights, and it can benchmark a model against a running provider afterward.

SetupDocsCommunityMaturity

Verdict Our llmfit build took 110 seconds and all 572 tests passed in 122 seconds, making it a credible first filter for local-model selection. Use its ranking to shorten a download list, then benchmark the winner on the exact runtime and workload you care about.

Full review →

03

opencode

anomalyco/opencode · ★ 208k ▲1.1k/7d · TypeScript

OpenCode is an open-source coding agent available in a terminal and a beta desktop app. It lets a chosen language model inspect code, edit files, run commands, call subagents, and use MCP tools under configurable permissions.

SetupDocsCommunityMaturity

Verdict OpenCode is an excellent trial for developers who care about provider choice, a polished terminal interface, and configurable agents. Its broad defaults put responsibility on the user to define a safe permission policy, and the source monorepo is far heavier than the binary install suggests.

Full review →

04

vllm

vllm-project/vllm · ★ 92k ▲418/7d · Python

vLLM is an inference engine and API server for running large language models on your own hardware. It handles batching, model memory, parallel execution, and common model APIs so teams do not have to build a serving layer around raw model code.

SetupDocsCommunityMaturity

Verdict Our vLLM build finished in 28 seconds, but the test suite stopped after 12 seconds because the development environment lacked `tblib`, so adopters get a credible serving engine with a less self-contained contributor setup. Use it when you need serious control over batching, memory, parallelism, and API compatibility across large models.

Full review →

05

Fabric

danielmiessler/Fabric · ★ 44k ▲49/7d · Go

Fabric is a command-line tool and prompt library for running repeatable AI tasks such as summarizing text, extracting ideas, analyzing claims, and processing video transcripts. It packages named Markdown prompts, called Patterns, so the same workflow can be piped through different local or hosted models.

SetupDocsCommunityMaturity

Verdict Our Fabric build took 126 seconds, then its test run ended with 32 passing packages and 1 failure after 26 seconds. It is still a strong choice for engineers who want readable prompts as versionable command-line tools across several providers, especially for text and transcript work.

Full review →

06

awesome-agent-skills

VoltAgent/awesome-agent-skills · ★ 34k ▲357/7d

Awesome Agent Skills is a large, organized directory of instruction packs for Claude Code, Codex, Gemini CLI, Cursor, and other coding agents. It helps developers find a skill for a specific job without trawling through GitHub, while keeping official team entries separate from community submissions.

SetupDocsCommunityMaturity

Verdict Awesome Agent Skills is a useful discovery index and a poor substitute for due diligence. Its official-team grouping, client path table, and active submission flow make it a strong place to start a search.

Full review →

07

freellmapi

tashfeenahmed/freellmapi · ★ 27k ▲1.1k/7d · TypeScript

FreeLLMAPI is a self-hosted gateway that puts many providers' free AI tiers and custom OpenAI-compatible services behind one API. It stores provider keys locally, tracks quotas, and switches models or providers when a request hits a limit, aiming at personal experimentation rather than production traffic.

SetupDocsCommunityMaturity

Verdict Our FreeLLMAPI run installed 778 packages in 32 seconds, then passed its 26-second build and 157-second test step with 0 audit findings. That makes it an unusually credible personal gateway for developers who already collect free-provider keys, but the project's own no-production warning should decide the boundary.

Full review →

08

copilot-sdk

github/copilot-sdk · ★ 10k ▲13/7d · Java

GitHub Copilot SDK embeds the agent runtime behind Copilot CLI into applications written in TypeScript, Python, Go, .NET, Java, or Rust. It handles sessions, planning, tools, file edits, custom agents, skills, hooks, and MCP connections through a CLI server controlled by the SDK.

SetupDocsCommunityMaturity

Verdict Our Rust SDK run installed 181 packages, built in 147 seconds, and passed all 260 tests in 187 seconds, giving this repository the strongest measured result in this batch. Use Copilot SDK when Copilot CLI behavior and 6 official language bindings reduce more work than the extra runtime process creates.

Full review →

09

llama-swap

mostlygeek/llama-swap · ★ 5.7k ▲53/7d · Go

llama-swap is a local reverse proxy that starts the right AI model server when a request names a model and stops conflicting servers when memory is needed elsewhere. It gives OpenAI- and Anthropic-compatible clients one address for text, embeddings, audio, images, reranking, and several local backends without keeping every model loaded at once.

SetupDocsCommunityMaturity

Verdict Our llama-swap build took 44 seconds and all 17 tested Go packages passed in 25 seconds, giving the proxy itself a clean baseline. Use it when local model variety exceeds available memory and your clients already speak OpenAI or Anthropic APIs.

Full review →

10

ux-ui-agent-skills

plugin87/ux-ui-agent-skills · ★ 1.5k · JavaScript

UX/UI Agent Skills is a reusable instruction and knowledge kit that gives Claude structured guidance for design systems, accessibility, component design, and interface code. It solves the consistency problem teams hit when an AI assistant produces attractive but disconnected screens without shared tokens, states, or quality gates.

SetupDocsCommunityMaturity

Verdict Our run installed in 6 seconds and completed its tests in 4 seconds. UX/UI Agent Skills is worth trying if Claude Code is already part of your front-end workflow and you want a serious shared vocabulary for tokens, accessibility, components, and review.

Full review →

11

ollama

ollama/ollama · ★ 181k ▲495/7d · Go

Ollama runs downloadable language and vision models on your own computer and exposes them through a command line and local REST API. It solves the fiddly parts of obtaining model files, choosing a runtime, and keeping a model available for apps that should not depend entirely on a hosted API.

SetupDocsCommunityMaturity

Verdict Our Ollama build took 116 seconds and all 60 tests passed, making it an unusually low-risk way to try local models if you accept that model choice and hardware still determine the experience. Use it for developer workstations, private prototypes, and integrations that benefit from one familiar local API.

Full review →

12

transformers

huggingface/transformers · ★ 166k ▲1.1k/7d · Python

Transformers is a Python library that defines and runs pretrained models for text, images, audio, video, and mixed-media tasks. It gives researchers and product teams common loading, preprocessing, inference, and training APIs across a very large model catalog.

SetupDocsCommunityMaturity

Verdict Transformers remains the first library to try when a published model has a Hugging Face implementation and you need direct control over inference or training. Its common APIs save real integration work, but the supported surface is enormous and the checkpoint is part of your dependency chain.

Full review →

13

langflow

langflow-ai/langflow · ★ 155k ▲295/7d · Python

Langflow is a visual builder for creating AI agents and workflows, testing them interactively, and exposing the finished flows as APIs or MCP tools. It helps developers connect models, retrieval, tools, and conversation logic without hand-writing every integration, while still allowing Python customization when the visual components are not enough.

SetupDocsCommunityMaturity

Verdict Our npm install succeeded in 5 seconds, but the repository exposed no npm build or test target. Langflow is worth trying if visual iteration, Python-level customization, and API or MCP publishing belong in one self-hosted tool.

Full review →

14

llama.cpp

ggml-org/llama.cpp · ★ 128k ▲578/7d · C++

llama.cpp runs language and vision-language models on local machines and servers through a C++ inference engine. It solves the practical problem of using quantized models across CPUs, Apple Silicon, and several GPU backends without depending on a hosted model API.

SetupDocsCommunityMaturity

Verdict Our llama.cpp build finished in 10 seconds, but the test command stopped with 2 collection errors and pip-audit found 6 known vulnerabilities, so the engine is easy to try and still demands a controlled deployment. Use it when local inference and hardware choice are worth owning model conversion, memory sizing, and backend tuning.

Full review →

15

TradingAgents

TauricResearch/TradingAgents · ★ 107k ▲3.1k/7d · Python

TradingAgents is a Python research framework that asks a group of language-model agents to analyze a ticker, argue the bull and bear cases, and return a five-level portfolio rating. It combines market, company, news, social, macro, and prediction-market inputs so a developer can study an agent-based research process without wiring the whole workflow. Execution is simulated; the project does not connect to a brokerage account.

SetupDocsCommunityMaturity

Verdict Our TradingAgents run installed 130 packages in 60 seconds, used 352 MB, and passed its build and tests, so the code is easy to trial as a research framework. Use it to inspect how language-model analysts debate a ticker and to prototype your own decision pipeline.

Full review →

16

unsloth

unslothai/unsloth · ★ 76k ▲214/7d · Python

Unsloth is a desktop app, web interface, and Python package for running and training local language, image, audio, embedding, and diffusion models. It tries to put model downloads, chat, fine-tuning, data preparation, export, agent connections, and local serving in one place.

SetupDocsCommunityMaturity

Verdict Our Unsloth run installed 45 packages in 40 seconds and passed all 393 executed tests, so its Python core deserves a trial from anyone who needs local fine-tuning or a combined model workspace. The clean build does not remove the cost of model weights, GPU-specific setup, or securing Studio's active tools.

Full review →

17

anything-llm

Mintplex-Labs/anything-llm · ★ 66k ▲246/7d · JavaScript

AnythingLLM is a desktop and self-hosted chat application that connects language models to your documents, tools, and shared workspaces. It gives people a ChatGPT-style interface while letting the operator choose local models, cloud providers, embedding engines, and vector storage.

SetupDocsCommunityMaturity

Verdict Our AnythingLLM run installed 288 packages in 52 seconds and passed 180 of 182 tests, but Jest still exited 1 with 34 failed suites. It is a sensible first trial for document chat when provider choice and a ready interface matter more than fine-grained retrieval control.

Full review →

18

LocalAI

mudler/LocalAI · ★ 49k ▲78/7d · Go

LocalAI is a self-hosted AI server that puts text, vision, speech, image, video, agents, and MCP tools behind familiar APIs. It solves the awkward job of operating several model engines by downloading separate backends only when a chosen model needs them.

SetupDocsCommunityMaturity

Verdict Our LocalAI run installed 1,246 packages and built in 11 seconds, but 2 of 34 tested Go packages failed, so it is a serious platform to trial rather than a server to trust untouched. Choose it when one private API must cover several model types and your team can own backend, model, and hardware choices.

Full review →

19

prompt-optimizer

linshenkx/prompt-optimizer · ★ 35k ▲397/7d · TypeScript

Prompt Optimizer is a TypeScript application for improving, testing, comparing, and saving prompts before you use them with an AI provider. It turns prompt tweaking from a copy-and-paste habit into a repeatable workspace, with text and image workflows available through a web app, desktop app, Chrome extension, Docker, or MCP.

SetupDocsCommunityMaturity

Verdict Our build finished in 54 seconds, but the test suite ended with 1 failure after 1,204 tests passed. Prompt Optimizer is worth trying if you want a polished, local-first workbench that joins prompt rewriting, comparison, reusable assets, image prompting, and MCP access.

Full review →

20

langfuse

langfuse/langfuse · ★ 35k ▲291/7d · TypeScript

Langfuse is an English-first platform for tracing, evaluating, and debugging applications that call language models. It stores model calls and surrounding application steps, then connects those traces to prompt versions, user feedback, datasets, costs, and evaluation results. Chinese, Japanese, and Korean README translations also exist.

SetupDocsCommunityMaturity

Verdict Our Langfuse install consumed 2,182 MB, then the web build stopped in 12 seconds on environment validation, so the source checkout is not a zero-config trial. Langfuse is a strong fit when a team will connect traces to prompt versions, datasets, and evaluations rather than merely count tokens.

Full review →

21

openclaude

Gitlawb/openclaude · ★ 33k ▲194/7d · TypeScript

OpenClaude is a terminal coding agent derived from Claude Code and adapted to work with many cloud APIs, gateways, Codex authentication, Gemini, GitHub Models, and local model servers. It keeps prompts, file and shell tools, subagents, MCP, slash commands, background jobs, and streaming output in one interface while letting the user choose the model provider.

SetupDocsCommunityMaturity

Verdict Our OpenClaude checkout installed in 47 seconds, built in 9 seconds, and completed its tests in 138 seconds, the only clean three-stage result among these seven reviews. Technically, it is a serious multi-provider coding-agent option with unusually broad tooling.

Full review →

22

ai

vercel/ai · ★ 27k ▲144/7d · TypeScript

AI SDK is a TypeScript toolkit for calling language-model providers, streaming responses, generating typed data, running tool loops, and wiring AI output into web interfaces. It gives applications one API across providers while still allowing direct provider packages when a feature needs vendor-specific access.

SetupDocsCommunityMaturity

Verdict Our AI SDK workspace install consumed 4,066 MB, and its build stopped after 67 of 91 tasks succeeded, so contributing to the monorepo is far heavier than installing the `ai` package. Use it when a TypeScript product needs streaming UI and provider choice behind one familiar API.

Full review →

23

opik

comet-ml/opik · ★ 22k ▲125/7d · Python

Opik records traces from language-model applications and gives teams a place to inspect calls, compare experiments, manage prompts, run evaluations, and watch production behavior. It can run as Comet's hosted service or as an Apache-2.0 self-hosted platform with Python, TypeScript, OpenTelemetry, and MCP entry points.

SetupDocsCommunityMaturity

Verdict Our Opik load-test subproject installed 108 packages and built in 28 seconds combined, but its test run timed out after 157 seconds while flushing an attachment trace. Opik is worth a trial when trace ownership, evaluations, and framework-neutral instrumentation need to live together.

Full review →

24

mcp-toolbox

googleapis/mcp-toolbox · ★ 16k ▲121/7d · Go

MCP Toolbox for Databases is a Go server that gives AI clients controlled access to databases through the Model Context Protocol. It supplies ready-made database tools and lets teams define narrower SQL, search, and administrative actions in YAML instead of handing an assistant a raw connection.

SetupDocsCommunityMaturity

Verdict Our mcp-toolbox run built in 261 seconds, but 49 of 452 tests failed because integration settings such as `VALKEY_ADDRESS` and `YUGABYTEDB_DATABASE` were missing. Use it when an agent needs governed access to several real databases and your team can own credentials, roles, and tool definitions.

Full review →

25

agents

livekit/agents · ★ 14k ▲134/7d · Python

LiveKit Agents is a Python framework for server-side voice and video agents that join real-time rooms as programmable participants. It connects speech recognition, language models, speech synthesis, turn detection, telephony, client apps, and MCP tools around a shared session and job runtime.

SetupDocsCommunityMaturity

Verdict Our LiveKit Agents install took 27 seconds and only 37 MB, but pytest stopped in 6 seconds because the `livekit` package could not be imported. The framework is a strong candidate when real-time media, telephony, dispatch, and model choice must work as one system.

Full review →

26

adk-go

google/adk-go · ★ 8.8k ▲32/7d · Go

Google's Agent Development Kit for Go is a code-first library for building, evaluating, orchestrating, and deploying AI agents. It is optimized for Gemini but exposes tools, multi-agent composition, model interfaces, sessions, and deployment pieces as Go packages.

SetupDocsCommunityMaturity

Verdict Our ADK Go run installed 289 packages, built successfully, and passed all 72 tests, so the core library deserves a trial for teams committed to Go. Use it when Gemini and cloud-native deployment are natural fits, and keep agent logic inside code you can test.

Full review →

27

notfair-plugin

nowork-studio/notfair-plugin · ★ 3.8k ▲71/7d · TypeScript

NotFair Plugin is a collection of 45 open-source marketing procedures for Claude Code, Codex, Cursor, Gemini CLI, and compatible agents. It gives an agent detailed instructions for SEO, analytics, and paid-media work, while a hosted MCP connection supplies live account data and approved write actions.

SetupDocsCommunityMaturity

Verdict Our NotFair Plugin run installed 56 packages in 34 seconds and all 273 executed pytest cases passed, making the repository a low-friction trial for teams willing to inspect agent instructions. The strongest reason to use it is the combination of readable marketing procedures and one OAuth-connected MCP for live accounts.

Full review →

28

stable-diffusion-webui

AUTOMATIC1111/stable-diffusion-webui · ★ 165k ▲80/7d · Python

Stable Diffusion WebUI is a local browser interface for generating and editing images with Stable Diffusion models. It gathers prompting, inpainting, upscaling, model switching, training utilities, and community extensions in one Gradio application.

SetupDocsCommunityMaturity

Verdict Stable Diffusion WebUI remains a capable workbench for people who know which models and settings they want. Its density and extension culture reward experimentation, while the Python, Torch, and GPU compatibility surface makes it a poor fit for hands-off deployment.

Full review →

29

dify

langgenius/dify · ★ 156k ▲550/7d · TypeScript

Dify is a visual platform for building applications around language models, retrieval, tools, and multi-step workflows. It gives product teams one place to design prompts, connect model providers, index documents, expose APIs, and inspect production runs.

SetupDocsCommunityMaturity

Verdict Our Dify install pulled 1,342 packages, used 1,408 MB, and exposed no root build or test target, so adoption starts with a large working surface and incomplete evidence from the default workspace command. Use it when a team needs a shared LLM application layer with workflows, retrieval, APIs, and run inspection.

Full review →

30

agency-agents

msitarzewski/agency-agents · ★ 153k ▲1.2k/7d · Shell

Agency Agents is a large collection of role-specific instruction files for coding assistants and other agent tools. It gives Claude Code, Codex, Cursor, Gemini CLI, and several peers reusable personas, workflows, deliverables, and review rules instead of asking users to write each role prompt themselves.

SetupDocsCommunityMaturity

Verdict Our Agency Agents run installed 35 packages and built in 17 seconds total, but it exposed no test target, so the collection is easy to install and hard to verify as a whole. Use it as a menu of prompts, selecting and editing a few roles that match your process.

Full review →

31

langchain

langchain-ai/langchain · ★ 146k ▲333/7d · Python

LangChain is a Python framework for connecting language models to tools, retrieval systems, structured outputs, and agent loops through shared interfaces. It helps teams change providers and compose common AI application parts without writing every adapter themselves.

SetupDocsCommunityMaturity

Verdict Our 79 MB LangChain core install finished in 21 seconds, but its configured test command stopped after 9 seconds on an unrecognized pytest argument, so adding it is easier than verifying its checkout. Use LangChain when model portability and a broad integration layer will save more code than the framework adds.

Full review →

32

hello-agents

datawhalechina/hello-agents · ★ 79k ▲947/7d · Python

Hello-Agents is a Chinese-first, 16-chapter course on building language-model agents; full English versions of the main README and chapters are available. It teaches agent patterns, memory, retrieval, protocols, training, evaluation, and multi-agent projects, with code that readers are expected to run and modify.

SetupDocsCommunityMaturity

Verdict Our DataAnalysisAgent install pulled 137 packages in 41 seconds and built in 4 seconds, but Hello-Agents exposed no repository test target, so treat it as a course to verify while learning. It is a strong free syllabus for a Python reader who wants 16 connected chapters and can work in Chinese or the English translation.

Full review →

33

crewAI

crewAIInc/crewAI · ★ 59k ▲286/7d · Python

CrewAI is a Python framework for arranging language-model agents into teams and placing their work inside event-driven flows. It tackles jobs where one prompt is too limited, especially when separate roles, tools, state, branching, or human review need to share one workflow.

SetupDocsCommunityMaturity

Verdict Our CrewAI checkout installed 34 packages in 29 seconds, but its build and test steps both failed, so the quick application path is more convincing than the contributor path we measured. Use CrewAI when role-based agents genuinely make a job easier to express, then wrap risky or repeatable work in Flows.

Full review →

34

llama_index

run-llama/llama_index · ★ 52k ▲62/7d · Python

LlamaIndex is a Python framework for building applications that retrieve private data, call models and tools, and run agent workflows. It saves teams from wiring every document reader, index, retriever, model provider, and vector store by hand.

SetupDocsCommunityMaturity

Verdict LlamaIndex is a strong first choice for a Python team whose main problem is connecting private data to models. Start with core and a short integration list, because the convenience package can hide how much provider-specific behavior you are adopting.

Full review →

35

dspy

stanfordnlp/dspy · ★ 38k ▲122/7d · Python

DSPy is a Python framework for expressing language-model programs as typed modules and then optimizing their instructions or examples against a metric. It replaces hand-edited prompt strings with reusable signatures, predictors, retrieval steps, agents, and optimizers.

SetupDocsCommunityMaturity

Verdict Our DSPy build finished in 10 seconds, yet 46 tests failed and 16 more hit collection or setup errors, so its ideas are easier to recommend than this exact checkout as a clean drop-in. Use DSPy when you can define a metric and want the framework to search for better instructions or demonstrations.

Full review →

36

sglang

sgl-project/sglang · ★ 36k ▲238/7d · Python

SGLang is an inference server for large language, vision-language, embedding, reward, and diffusion models. It turns model weights into OpenAI-compatible and native HTTP APIs, with scheduling, caching, parallel execution, and quantization for anything from one accelerator to a distributed cluster.

SetupDocsCommunityMaturity

Verdict Our Rust-subproject run built in 178 seconds and passed all 580 tests, which makes SGLang credible for an experienced inference team despite its size. Use it when accelerator utilization, model breadth, and distributed serving justify a dedicated platform.

Full review →

37

SillyTavern

SillyTavern/SillyTavern · ★ 33k ▲183/7d · JavaScript

SillyTavern is a locally installed chat interface for talking to many hosted or local language models through one highly configurable screen. It adds character cards, lorebooks, group chats, prompt controls, image generation, and speech integrations for people who find ordinary assistant chat boxes too restrictive.

SetupDocsCommunityMaturity

Verdict Our SillyTavern install pulled 785 packages and 400 MB in 19 seconds, while npm audit reported 45 known vulnerabilities, including 1 critical. Use it when character chat and fine prompt control justify owning a large, fast-moving local application.

Full review →

38

agent-skills

vercel-labs/agent-skills · ★ 31k ▲177/7d · JavaScript

Vercel Agent Skills is an English-language collection of instruction packages and helper scripts for coding agents. Its documented skills teach agents how to review React, React Native, interfaces, and writing, audit Vercel projects, work with view transitions and composition, or deploy an application to Vercel.

SetupDocsCommunityMaturity

Verdict Our Agent Skills install added only 2 packages in 8 seconds and npm audit found 0 known vulnerabilities, but the repository exposed neither build nor test targets. Install individual skills only when their instructions match the task, then review every script, remote fetch, credential requirement, and upload path.

Full review →

39

book-to-skill

virgiliojr94/book-to-skill · ★ 31k ▲824/7d · Python

book-to-skill turns a book or document folder into a set of instructions and topic files that a coding agent can consult later. It extracts the source locally, then guides Claude Code, Codex, Copilot CLI, Amp, or Hermes Agent through writing chapter notes, a glossary, patterns, and a quick-reference file.

SetupDocsCommunityMaturity

Verdict Our book-to-skill run installed 161 packages and consumed 6,075 MB, then passed all 627 tests in 15 seconds. Use it when the same owned documents keep returning in agent-assisted work and you are prepared to review the resulting skill as edited notes.

Full review →

40

claude-code-templates

davila7/claude-code-templates · ★ 31k ▲162/7d · Python

Claude Code Templates is a catalog and CLI for installing agents, slash commands, settings, hooks, skills, MCP integrations, and plugins into Claude Code. It also includes local dashboards for session analytics, chat monitoring, health checks, skills, and plugin management.

SetupDocsCommunityMaturity

Verdict Our Claude Code Templates install took 11 seconds and its configured tests returned success, but npm audit found 12 vulnerabilities and the test script runs no behavioral tests. Use the project as a discovery catalog whose individual components must earn trust before installation, especially hooks and MCP servers.

Full review →

41

openai-agents-python

openai/openai-agents-python · ★ 29k ▲129/7d · Python

The OpenAI Agents SDK is a Python framework for running language-model agents with tools, handoffs, guardrails, session memory, human approvals, and tracing. It gives developers one runner for text, long-running sandbox work, realtime sessions, and voice pipelines instead of making each application build those control loops itself.

SetupDocsCommunityMaturity

Verdict Our install took 15 seconds, but pytest reported 135 failures and 65 collection/setup errors, so the OpenAI Agents SDK is easy to try and hard to approve from the README path alone. Use it for a Python agent product that needs OpenAI Responses, tools, handoffs, approvals, and tracing under one runner.

Full review →

42

deepagents

langchain-ai/deepagents · ★ 29k ▲156/7d · Python

Deep Agents is LangChain's opinionated Python harness for agents that work across long, multi-step tasks. It bundles planning, file and shell tools, subagents, context compression, memory, human approval, skills, and MCP tools on top of LangChain and LangGraph, while letting developers replace those pieces.

SetupDocsCommunityMaturity

Verdict Our Deep Agents install added 82 packages and built in 30 seconds combined, but pytest collected 0 tests and exited 4 on an unknown asyncio option. The framework is a strong candidate for teams already committed to LangChain that want files, delegation, memory, and MCP in one harness, provided every dangerous tool runs inside a real sandbox.

Full review →

43

promptfoo

promptfoo/promptfoo · ★ 25k ▲150/7d · TypeScript

Promptfoo is a command-line tool and library for testing prompts, models, agents, and retrieval-based AI applications. It lets teams compare outputs, turn expected behavior into repeatable checks, and probe an AI system for security failures before release.

SetupDocsCommunityMaturity

Verdict Our Promptfoo checkout installed 2,415 packages and passed 23,460 tests, but its source build failed after 79 seconds, so it fits teams willing to own a large evaluation stack better than developers seeking a light dependency. The 30 audit findings, including 27 high-severity vulnerabilities, require review before putting it in a sensitive CI environment.

Full review →

44

hindsight

vectorize-io/hindsight · ★ 24k ▲330/7d · Python

Hindsight is a memory server for AI agents that stores facts and experiences, retrieves relevant material, and maintains synthesized knowledge as new information arrives. It can sit behind agent frameworks, coding assistants, an MCP client, or a custom application through Python, Node.js, Go, CLI, and REST clients.

SetupDocsCommunityMaturity

Verdict Our Hindsight checkout consumed 1,704 MB for 2,141 npm packages, reported 23 audit findings, and offered no measured build or test target. Trial it when long-lived agent memory is a product requirement and you can evaluate retrieval quality with your own conversations.

Full review →

45

verl

verl-project/verl · ★ 23k ▲66/7d · Python

verl is a Python framework for reinforcing and fine-tuning large language models after their initial training. It coordinates training backends, rollout engines, rewards, model placement, and distributed workers so research teams can run PPO, GRPO, and related methods on their own GPU infrastructure.

SetupDocsCommunityMaturity

Verdict Our verl install consumed 6,096 MB and its tests ended with exit 137 at 9% after 74 seconds, so adoption should be treated as an infrastructure project even though the build passed. Choose it when your team needs distributed RL across several training and rollout backends and can own the GPU, Ray, and version matrix.

Full review →

46

prime-agent

PrimeIntellect-ai/prime-agent · ★ 21k ▲362/7d · TypeScript

Prime Agent is a terminal coding and research agent built for jobs that continue across long sessions. It gives the model a persistent Python environment, real subagents, background execution, schedules, and a supplemental memory and skill system that can be refined and rolled back.

SetupDocsCommunityMaturity

Verdict Our Prime Agent test command ran for 567 seconds and exited 1 despite reporting 750 passed and 0 failed, while npm audit found 2 high and 2 moderate vulnerabilities. That is too ambiguous for unattended rollout, but the 13-second build and unusually explicit trust warning make a controlled research trial reasonable.

Full review →

47

web-llm

mlc-ai/web-llm · ★ 19k ▲43/7d · TypeScript

WebLLM runs supported language models inside a web browser with WebGPU, so prompts and generated text can stay on the user's device after model assets arrive. Its TypeScript API resembles OpenAI chat completions and includes streaming, JSON output, workers, browser caching, and custom MLC model support. It solves the server-cost and data-routing problem for web apps whose users have suitable hardware.

SetupDocsCommunityMaturity

Verdict Our run of WebLLM installed 428 packages, built in 20 seconds, and passed all 209 tests, making it the most convincing browser LLM package in this set. Use it when local inference is a product requirement and you can publish a tested browser, GPU, model, and cache matrix.

Full review →

48

claude-code-system-prompts

Piebald-AI/claude-code-system-prompts · ★ 13k ▲48/7d · JavaScript

Claude Code System Prompts is an unofficial, versioned archive of prompt text extracted from Claude Code's compiled npm package. It separates hundreds of system instructions, tool descriptions, agent prompts, reminders, skills, and data blocks into searchable Markdown files so researchers and prompt authors can see what changed.

SetupDocsCommunityMaturity

Verdict This is the best kind of unofficial reference: current, granular, diffable, and honest in its repository guidance about not being the real source code. Use it to investigate behavior and track prompt evolution, not as a vendor contract or a drop-in prompt library.

Full review →

49

claude-skills

Jeffallan/claude-skills · ★ 12k · Python

Claude Skills is a Claude Code plugin containing 67 specialist instruction sets for software work, plus nine commands for running projects through Jira and Confluence. It gives the coding agent reusable methods and reference material for particular languages, frameworks, security reviews, testing, and operations.

SetupDocsCommunityMaturity

Verdict Our claude-skills checkout installed 35 Python packages in 22 seconds and built in 6 seconds, but it offered no test target for judging what the skills make an agent produce. Use it as an editable reference library if your Claude Code work crosses several stacks and you will review every consequential action.

Full review →

50

rig

0xPlaygrounds/rig · ★ 8.7k ▲51/7d · Rust

Rig is a Rust library for adding language models, tools, memory, and retrieval to an application without writing a separate client for every provider. It supplies shared interfaces and an agent runtime, while leaving the surrounding product, deployment, and model accounts to you.

SetupDocsCommunityMaturity

Verdict Our Rig build succeeded in 242 seconds, but cargo test finished with 3,870 of 3,872 tests passing, so adoption should begin behind an internal wrapper. Use it when Rust is fixed, multiple providers are likely, and the team can budget for v0.x migrations.

Full review →

51

bifrost

maximhq/bifrost · ★ 8.1k ▲130/7d · Go

Bifrost is a self-hosted gateway that gives applications one API for more than 23 model providers. It centralizes credentials, request translation, retries, fallbacks, budgets, logs, plugins, and MCP tools so each application does not have to implement those controls separately.

SetupDocsCommunityMaturity

Verdict Our Bifrost run installed 138 packages and built in 96 seconds, but 1 of 38 Go test packages failed after 330 seconds. Shortlist it when provider routing, policy, and MCP management have become a shared platform problem, then run contract tests against every provider feature you use.

Full review →

52

awesome-free-llm-apis

mnfst/awesome-free-llm-apis · ★ 7.7k ▲212/7d · JavaScript

Awesome Free LLM APIs is a curated English-language catalog of hosted text-model APIs with recurring free access. It puts signup links, API base URLs, model names, context sizes, output limits, modalities, and quotas in one place, so developers can shortlist providers without opening a dozen pricing pages.

SetupDocsCommunityMaturity

Verdict Awesome Free LLM APIs is an excellent first stop and a poor final authority. Its structured tables, source data, verification script, and blunt notes about privacy and signup restrictions beat most link collections.

Full review →

53

engram

Gentleman-Programming/engram · ★ 6.7k ▲127/7d · Go

Engram gives coding agents a persistent memory store backed by SQLite and FTS5. A single Go binary exposes that memory through MCP, a CLI, a terminal interface, and an optional HTTP or cloud layer, so work from one session can be found in the next.

SetupDocsCommunityMaturity

Verdict Our Engram run built in 40 seconds, but 2 of 22 test results failed after a 476-second suite, including one plugin check that required `jq` and `curl`. Use it when local SQLite memory shared across several coding agents is worth explicit save and retrieval habits.

Full review →

54

Claude-Red

SnailSploit/Claude-Red · ★ 5.8k · Python

Claude-Red is a library of ready-made instruction files that guide Claude through authorized offensive-security work. It gives operators structured methods for 78 topics, so they do not have to rebuild a testing checklist for every engagement.

SetupDocsCommunityMaturity

Verdict Our run installed 35 packages in 54 seconds and built in 1 second, but it had no test target to run. Claude-Red is worth trying if you already have authorization controls and want a broad, neatly indexed set of Claude Code playbooks.

Full review →

55

flash-linear-attention

fla-org/flash-linear-attention · ★ 5.8k ▲16/7d · Python

Flash Linear Attention is a Python library of GPU-oriented kernels, layers, and model definitions for linear attention, sparse attention, state-space models, and hybrid language models. It gives researchers a shared implementation base for trying newer sequence architectures without writing every Triton kernel and backward pass themselves.

SetupDocsCommunityMaturity

Verdict Our Flash Linear Attention build finished in 6 seconds, but its tests stopped before collection because the 57-package base install did not include Torch. Researchers should use it when they need several emerging sequence architectures under one well-tested, hardware-aware codebase and can install the correct backend extra.

Full review →

56

agentgateway

agentgateway/agentgateway · ★ 4.9k ▲79/7d · Rust

Agentgateway is a proxy for traffic between agents, language models, MCP tool servers, and other agents using A2A. It centralizes routing, credentials, policy, guardrails, spending controls, and telemetry, with standalone and Kubernetes deployment modes.

SetupDocsCommunityMaturity

Verdict Our Agentgateway build took 362 seconds, then all 78 measured tests passed in 193 seconds, which is a credible result for a large gateway codebase. Use it when LLM, MCP, and A2A traffic genuinely need one policy and telemetry layer, especially on Kubernetes.

Full review →

57

cc-skills-golang

samber/cc-skills-golang · ★ 3.3k ▲39/7d · Go

cc-skills-golang is a collection of reusable instructions that teach coding agents how to handle Go style, architecture, testing, security, performance, and popular libraries. It gives Claude Code and other Agent Skills-compatible assistants a reviewed Go playbook instead of relying only on the model's general training.

SetupDocsCommunityMaturity

Verdict cc-skills-golang is a valuable accelerator for Go work because it packages specific, inspectable advice instead of promising model magic. Its breadth and cross-client installation are impressive, and active corrections are a healthy sign.

Full review →

58

agentdock

uvwt/agentdock · ★ 972 ▲141/7d · Go

AgentDock is an MCP runtime that lets an AI client operate files, shells, Git repositories, browsers, and long-running tasks on machines you control. It supplies the connection and permission layer between clients such as ChatGPT, Claude, or Codex and local computers, servers, or containers; it does not supply a chat interface or model inference.

SetupDocsCommunityMaturity

Verdict Our AgentDock run built in 31 seconds and passed all 82 tests, so the checked-out runtime is a credible candidate for an isolated, least-privilege host. Use it when one trusted AI client genuinely needs to work across several machines and you can narrow each service account or container.

Full review →

59

open-webui

open-webui/open-webui · ★ 152k ▲630/7d · Python

Open WebUI is a self-hosted chat and workspace interface for local models, OpenAI-compatible services, documents, tools, and shared teams. It gives people one browser-based home for several AI backends instead of tying every conversation to a provider's own app.

SetupDocsCommunityMaturity

Verdict Our Open WebUI build passed after 180 seconds, but the 1,263-package install carried 29 known vulnerabilities and no test target was available. Use it when a shared, self-hosted AI workspace is worth owning a large application and its upgrade work.

Full review →

60

awesome-llm-apps

Shubhamsaboo/awesome-llm-apps · ★ 139k ▲1.3k/7d · Python

Awesome LLM Apps is a large collection of runnable Python examples for agents, RAG, voice tools, MCP integrations, and other LLM projects. It solves the blank-page problem for developers who learn faster from working code than from framework documentation, though each folder should be judged as its own small project.

SetupDocsCommunityMaturity

Verdict Our tarot-folder run installed 93 packages in 49 seconds and built in 4 seconds, but it had no test target, so the result supports learning and inspection rather than production trust. Use Awesome LLM Apps as a code shelf, one folder at a time.

Full review →

61

ComfyUI

comfyanonymous/ComfyUI · ★ 134k ▲945/7d · Python

ComfyUI is a local visual workspace for building image, video, audio, 3D, and text generation pipelines as node graphs. It solves the problem of reproducing a complicated generative workflow without hiding models, parameters, or processing steps behind a single prompt box.

SetupDocsCommunityMaturity

Verdict Our ComfyUI checkout installed 124 packages and used 5,903 MB, then 44 tests failed and 65 hit collection or setup errors, so adopting it means accepting a large environment whose tested commit was not clean on our box. Use it when graph-level control, repeatable visual pipelines, and broad model support justify the learning and maintenance.

Full review →

62

OpenHands

OpenHands/OpenHands · ★ 88k ▲645/7d · TypeScript

OpenHands now centers on Agent Canvas, a self-hosted control panel for running coding agents and scheduled automations. It can operate the OpenHands agent, Claude Code, Codex, Gemini, or another Agent Client Protocol backend from one interface, on a laptop or remote server.

SetupDocsCommunityMaturity

Verdict Our OpenHands run installed 1,998 packages and used 1,043 MB, then its build and tests passed, so this is a serious control plane with a heavy local footprint. The current Canvas is ambitious and active, but its beta label and wide security boundary rule out a casual public deployment.

Full review →

63

OpenHands

All-Hands-AI/OpenHands · ★ 88k ▲645/7d · TypeScript

OpenHands now centers on Agent Canvas, a self-hosted control panel for running coding agents and scheduled automations. It can operate the OpenHands agent, Claude Code, Codex, Gemini, or another Agent Client Protocol backend from one interface, on a laptop or remote server.

SetupDocsCommunityMaturity

Verdict Our OpenHands run installed 1,998 packages and used 1,043 MB, then its build and tests passed, so this is a serious control plane with a heavy local footprint. The current Canvas is ambitious and active, but its beta label and wide security boundary rule out a casual public deployment.

Full review →

64

lobehub

lobehub/lobe-chat · ★ 83k ▲130/7d · TypeScript

LobeHub, formerly published at lobehub/lobe-chat, is a web and desktop workspace for creating agents, grouping them into teams, scheduling work, and connecting them to model providers and tools. It is meant to replace a pile of separate chat tabs with projects, shared context, editable memory, and agents that can keep working after you leave.

SetupDocsCommunityMaturity

Verdict Our npm install failed after 303 seconds on `workspace:*`, so LobeHub is a poor choice for teams unwilling to follow its pnpm-based monorepo workflow. It is still one of the more ambitious self-hosted agent workspaces for teams that want projects, schedules, model choice, MCP tools, and messaging in one product.

Full review →

65

cline

cline/cline · ★ 68k ▲511/7d · TypeScript

Cline is an open-source coding agent that works in a terminal, VS Code, JetBrains IDEs, or through a Node.js SDK. It can inspect a repository, edit files, run commands, call MCP tools, and keep long tasks moving while giving the user a choice between approvals and unattended execution.

SetupDocsCommunityMaturity

Verdict Our Cline run installed 2,435 packages and used 3,838 MB, then its test command failed after 500 seconds, so this is a capable agent platform with a costly contributor setup. It is worth trying if you need the same agent across an IDE, CLI, CI, and custom SDK work, especially when model choice matters.

Full review →

66

openinterpreter

openinterpreter/openinterpreter · ★ 68k ▲54/7d · Rust

Open Interpreter is now a Rust coding agent built from OpenAI's Codex codebase, with selectable harnesses for lower-cost and local models. It runs commands on your computer, connects to editors through ACP, reads shared agent instructions and skills, and can speak the Codex exec protocol.

SetupDocsCommunityMaturity

Verdict Our Open Interpreter checkout installed 1 pnpm package in 40 seconds, but exposed no root build or test target, so that clean install says almost nothing about whether the Rust agent works. Try it if comparing low-cost model harnesses, ACP, MCP, and Codex-compatible execution is the actual job.

Full review →

67

gpt4free

xtekky/gpt4free · ★ 67k ▲6/7d · Python

GPT4Free, packaged as `g4f`, puts many hosted and local model adapters behind a Python client, web interface, OpenAI-compatible server, and MCP server. It can route text and media requests through providers with different login methods, but it does not make those providers equally reliable or grant permission to use their services.

SetupDocsCommunityMaturity

Verdict Our gpt4free install pulled 137 packages, occupied 507 MB, and built in 4 seconds, but it offered no test target to check the 76,608-line source tree. Use it for research and disposable experiments when provider churn is part of the work.

Full review →

68

claude-cookbooks

anthropics/claude-cookbooks · ★ 53k ▲120/7d · Jupyter Notebook

The Claude Cookbooks repository is Anthropic's official collection of example Jupyter notebooks and markdown guides for the Claude API. It shows developers how to wire up text classification, retrieval augmented generation, tool use, vision, prompt caching, and cost tuning with copy-able Python snippets.

SetupDocsCommunityMaturity

Verdict Our sandbox run installed the repo cleanly in 76 seconds but still left 163 of 1,731 pytest cases failing; that makes this a browsing library, not a buildable product. Use it to lift concrete Claude patterns into your own code, then keep your own tests green.

Full review →

69

LibreChat

danny-avila/LibreChat · ★ 44k ▲1.2k/7d · TypeScript

LibreChat is a self-hosted chat and agent workspace that puts many hosted and local AI providers behind one web interface. It gives teams their own accounts, conversations, agents, MCP tools, file workflows, and admin controls instead of making everyone work from separate vendor apps.

SetupDocsCommunityMaturity

Verdict Our LibreChat install completed in 78 seconds and occupied 2,295 MB, but the source build failed 20 seconds later in `librechat-data-provider`, so we would not adopt it without pinning a revision and reproducing the build in CI. It is a strong fit for a technical team that wants one self-hosted home for providers, agents, MCP, and user administration.

Full review →

70

LightRAG

HKUDS/LightRAG · ★ 40k ▲126/7d · Python

LightRAG turns a document collection into both a semantic search index and a map of entities and relationships, then uses both to find context for language-model answers. It is for questions that depend on connections spread across several passages, where ordinary similarity search can miss the relationship.

SetupDocsCommunityMaturity

Verdict Our LightRAG build installed 85 packages and completed, but pytest stopped with 60 collection or setup errors, so this is not a clean drop-in for a fresh Debian box. Pilot it when cross-document relationships justify running extraction models and four storage roles.

Full review →

71

colibri

JustVugg/colibri · ★ 35k ▲7.3k/7d · C

Colibri is a compact C inference engine that streams mixture-of-experts model weights across storage, system memory, and optional GPU memory. It is built for running unusually large open-weight models on hardware that cannot keep every expert in fast memory, with chat, server, web, and planning commands around the engine.

SetupDocsCommunityMaturity

Verdict Our Colibri run installed 34 Python packages, built in 10 seconds, and found 0 known vulnerabilities, but it ran no tests and loaded no model weights. Colibri is worth studying if expert streaming itself is the problem you want to solve and you can dedicate hundreds of gigabytes of storage to the experiment.

Full review →

72

zeroclaw

zeroclaw-labs/zeroclaw · ★ 33k ▲90/7d · Rust

ZeroClaw is a self-hosted personal assistant that connects language models to chat channels, local tools, browser tasks, memory, and scheduled procedures. One Rust program can answer through services such as Telegram or Matrix, use hosted or local models, and ask for approval before riskier actions.

SetupDocsCommunityMaturity

Verdict Our ZeroClaw build took 431 seconds and 407 of 408 tests passed; the lone failure came from a release script calling an absent `jq`, so serious self-hosters have a promising build with one concrete packaging prerequisite to fix. Choose it when channel reach, provider choice, and approval-gated procedures justify owning a 912,596-line agent runtime.

Full review →

73

heretic

p-e-w/heretic · ★ 32k ▲426/7d · Python

Heretic automatically modifies transformer-based language models to reduce safety refusals without conventional post-training. It searches for directional-ablation settings that suppress refusals while trying to keep the modified model close to the original, giving researchers and local-model builders a repeatable alternative to tuning those settings by hand.

SetupDocsCommunityMaturity

Verdict Our build finished in 7 seconds and all 4 tests passed, but pip-audit still found 2 known vulnerabilities. Heretic is one of the clearer options for automating abliteration, especially when reproducibility and parameter search matter.

Full review →

74

semantic-kernel

microsoft/semantic-kernel · ★ 29k ▲15/7d · C#

Semantic Kernel is Microsoft's software kit for connecting applications to language models and organizing assistants, tools, memory, and multi-agent workflows. It supports Python, .NET, and Java, but Microsoft now identifies Microsoft Agent Framework 1.0 as its production-ready successor, so new adopters must weigh current capability against a future migration.

SetupDocsCommunityMaturity

Verdict Our Python install completed in 43 seconds and build in 9 seconds, but tests exited 1 after 49 seconds. Semantic Kernel remains a capable, active option for existing deployments and teams that specifically need its provider, plugin, workflow, and multi-language surface.

Full review →

75

distilly

titanwings/colleague-skill · ★ 25k ▲174/7d · Python

Colleague Skill has been renamed Distilly. It turns messages, documents, interviews, and public sources into a portable Person Profile that Claude Code, Codex, and other skill-aware agents can load to imitate a person's working preferences or communication patterns without claiming to clone that person.

SetupDocsCommunityMaturity

Verdict Our Distilly run installed 42 packages, built in 5 seconds, and passed all 35 tests with 0 audit findings, so the code is easy to trial. Use it for a consented work profile when a portable Claude Code or Codex skill is more useful than a full memory service.

Full review →

76

knowledge-work-plugins

anthropics/knowledge-work-plugins · ★ 24k · Python

Knowledge Work Plugins is Anthropic's collection of 11 role-based plugin packs for Claude Cowork and Claude Code. Each pack combines instructions, slash commands, sub-agents, and MCP connectors so Claude can follow a repeatable workflow for jobs such as sales, finance, support, data analysis, or bio-research.

SetupDocsCommunityMaturity

Verdict Our run of the nested Allotrope skill installed 35 packages in 18 seconds and built in 4 seconds, but the repository offered no test target. Use this collection as editable Claude scaffolding when your team can inspect prompts, authorize each MCP connection, and review the output.

Full review →

77

notebooklm-py

teng-lin/notebooklm-py · ★ 19k ▲82/7d · Python

notebooklm-py is an unofficial Python client for Google's NotebookLM, now branded Gemini Notebook. It lets applications, shell scripts, MCP clients, and coding agents create notebooks, add sources, ask cited questions, generate media or study material, and download the results.

SetupDocsCommunityMaturity

Verdict Our notebooklm-py suite reached 76% before our 900-second cap, after a 31-second install and 7-second build, so the package is easy to obtain but slow to verify fully. Use it for personal research automation, prototypes, and agent tools where NotebookLM's cited answers and generated artifacts save real work.

Full review →

78

Megatron-LM

NVIDIA/Megatron-LM · ★ 18k ▲56/7d · Python

Megatron-LM is NVIDIA's reference framework for training very large transformer models across many GPUs. Its Megatron Core library supplies reusable model components and several ways to split training work across devices, while the repository adds scripts and examples for researchers running full training jobs.

SetupDocsCommunityMaturity

Verdict Megatron-LM belongs on the shortlist for teams with NVIDIA clusters and engineers who understand distributed training failures. Megatron Core offers unusual control over parallelism and model construction, but that control carries a large configuration and operations bill.

Full review →

79

agent-framework

microsoft/agent-framework · ★ 14k ▲70/7d · Python

Microsoft Agent Framework is a Python and .NET toolkit for building AI agents and joining them into stateful workflows. It supplies the machinery around model calls, including tools, middleware, checkpoints, streaming, human approval, tracing, and hosting patterns.

SetupDocsCommunityMaturity

Verdict Our Agent Framework install pulled 233 packages and occupied 912 MB, then the full test run failed after 72 seconds, so adopting it requires a team willing to verify its exact workflow path. It is a strong candidate when Python and .NET parity, Foundry integration, checkpoints, or multi-agent graphs are firm requirements.

Full review →

80

adk-samples

google/adk-samples · ★ 10k ▲29/7d · Python

google/adk-samples is a collection of small agents and larger recipes built with Google's Agent Development Kit. Its curated `core` examples teach patterns such as OAuth, memory, safety, research, and retrieval, while `contrib` holds community examples for particular workflows.

SetupDocsCommunityMaturity

Verdict Our adk-samples run built in 5 seconds and passed 900 tests, but the suite still exited 1 because 9 collection and setup checks could not import `tomlkit`. Use the repository as a recipe shelf for learning ADK or seeding a prototype, especially when a core example matches the exact pattern you need.

Full review →

81

no-ai-slop

petergyang/no-ai-slop · ★ 10k ▲1.7k/7d · Python

No AI Slop is an English-language editing skill for ChatGPT, Claude Code, Codex, and other coding agents. It asks the agent to remove more than 20 recurring writing habits while preserving the writer's wording and voice, and it has a separate mode that flags patterns without claiming to detect AI authorship.

SetupDocsCommunityMaturity

Verdict Our no-ai-slop run installed 35 packages in 39 seconds and built the plugin in 7 seconds, but the repository provided no behavioral test target. Use it as a compact editing brief when a human will compare the revision with the source.

Full review →

82

ai-memory

akitaonrails/ai-memory · ★ 7.0k ▲711/7d · Rust

ai-memory is a self-hosted service that records coding-agent activity, compiles it into a searchable Markdown wiki, and hands a bounded summary to a later session. It is built for developers moving work between Claude Code, Codex, and other agent clients who are tired of repeating decisions and failed approaches.

SetupDocsCommunityMaturity

Verdict Our ai-memory run passed 764 of 767 tests after a 238-second build, while the packaging failures included a missing Docker executable and a refused checksum update. Try it locally if cross-agent context loss costs you time and you can inspect the retained record.

Full review →

83

semantic-router

vllm-project/semantic-router · ★ 5.8k ▲148/7d · Go

vLLM Semantic Router sits between applications and several language-model backends, then chooses or combines model paths from request signals and policy. It gives platform teams one place to encode decisions about cost, latency, privacy, safety, model capability, caching, and multi-model workflows.

SetupDocsCommunityMaturity

Verdict Our Semantic Router run installed 78 packages and built successfully, but its test command ended with 73 passes and 6 collection or setup errors. That is acceptable evidence for an evaluation by an experienced inference team, not a clean release check.

Full review →

84

axonhub

looplj/axonhub · ★ 5.2k ▲29/7d · Go

AxonHub's primary README is Chinese, although English and Japanese documentation links are provided. It is a self-hosted AI gateway that lets applications keep using an OpenAI or Anthropic SDK while routing requests to different supported providers. It also puts tracing, access control, quotas, load balancing, and cost tracking in one control layer.

SetupDocsCommunityMaturity

Verdict Our build completed in 183 seconds, and all 78 Go tests passed. AxonHub is credible enough for a serious staging trial, particularly when SDK compatibility, tracing, RBAC, and cost controls must live together.

Full review →

85

OB1

NateBJones-Projects/OB1 · ★ 4.6k ▲13/7d · TypeScript

Open Brain is a personal memory backend that lets several AI clients save and retrieve the same information. It combines a Supabase database, semantic search, an MCP server, and optional recipes so your notes are no longer trapped inside one assistant.

SetupDocsCommunityMaturity

Verdict Our OB1 run installed 108 packages in 20 seconds and passed its test command in 8 seconds, so the code is cheap to inspect even though the green test does not establish the deployed server contract. Use it if you want shared personal memory across AI clients and accept Supabase plus OpenRouter as part of the design.

Full review →

86

embabel-agent

embabel/embabel-agent · ★ 4.5k ▲10/7d · Kotlin

Embabel Agent is a Kotlin and Java framework for building AI agents as typed goals, actions, conditions, and domain objects. Its planner chooses a route through the actions at runtime, letting Spring teams mix ordinary application code with model calls without encoding every flow as a fixed sequence.

SetupDocsCommunityMaturity

Verdict Our Embabel run passed 2,025 of 2,123 tests, skipped 98, and failed none, which makes version 1.5.1 a credible trial for a serious Spring team. Use it when typed domain objects and runtime planning solve a flow problem you can already name.

Full review →

87

gemini-skills

google-gemini/gemini-skills · ★ 4.2k · Python

Gemini Skills is a set of instruction files and helper scripts that teaches coding agents how to use current Gemini APIs and SDKs. It covers ordinary API work, live audio and video sessions, and Gemini Omni video generation so an agent is less likely to produce code from outdated model knowledge.

SetupDocsCommunityMaturity

Verdict Our lab could not execute commit 80dd31d because the Python repository had no supported entry point and no Dockerfile, so there is no install or test outcome to back the instructions. The skills are worth adding for developers who already chose Gemini and will pin, inspect, and refresh them.

Full review →

88

marin

marin-community/marin · ★ 3.7k ▲88/7d · Python

Marin is an open Python platform for researching and training foundation models, with data preparation, tokenization, training, evaluation, and cluster execution in one project. It records experiment code, checkpoints, reports, decisions, and failed runs so another research group can inspect how a model was made.

SetupDocsCommunityMaturity

Verdict Our Marin run installed 225 packages and used 1,954 MB, yet pytest stopped after 13 seconds on its configured arguments, so this platform belongs with research teams prepared to own the environment. Marin is worth studying when the experiment record matters as much as the checkpoint.

Full review →

89

docker-agent

docker/docker-agent · ★ 3.3k ▲11/7d · Go

Docker Agent is a Go runtime and Docker CLI plugin for defining AI agents in YAML, giving them tools, and letting them delegate work to other agents. It connects to cloud or local models, accepts MCP tools, can index documents for retrieval, and packages agent configurations in OCI registries.

SetupDocsCommunityMaturity

Verdict Our Docker Agent run built in 213 seconds, then 267 of 271 tests passed before the suite exited with 4 failures, so teams should trial it behind explicit safety and cost limits before standardizing on it. It is a strong fit when YAML, Docker distribution, MCP, and several model providers belong in the same operating layer.

Full review →

90

llm-for-zotero

yilewang/llm-for-zotero · ★ 3.0k ▲92/7d · TypeScript

llm-for-zotero puts an AI assistant inside Zotero for asking questions about papers, comparing sources, inspecting figures, and saving research notes. Its optional agent can also search and change a library, while MCP, Codex, Claude Code, WebChat, and local model connections let users choose where the model runs.

SetupDocsCommunityMaturity

Verdict Our llm-for-zotero run installed 432 packages and built successfully, but 1 of 5,358 tests failed and npm audit found 19 known vulnerabilities. It is worth trying if Zotero is already the center of your research and cited paper chat, notes, and controlled library actions belong in one interface.

Full review →

91

agent-router

theagentrouter/agent-router · ★ 2.1k ▲74/7d · Go

Agent Router is the new name for Envoy AI Gateway, an open-source control plane that puts one OpenAI-compatible address in front of hosted models, self-hosted inference, and MCP servers. It centralizes provider credentials, routing, quotas, fallback, and usage records while Envoy handles the network traffic.

SetupDocsCommunityMaturity

Verdict Our Agent Router build completed in 317 seconds, but its 653-second test run failed 16 of 28 reported tests, so source adopters need Kubernetes upgrade coverage of their own. Use it when AI routing has become a platform responsibility and Envoy Gateway is already acceptable infrastructure.

Full review →

92

agent-router

envoyproxy/ai-gateway · ★ 2.1k ▲58/7d · Go

Envoy AI Gateway is now Agent Router, an Envoy-based control plane that gives applications one OpenAI-compatible route to hosted models, self-hosted inference, and MCP servers. It centralizes credentials, routing, quotas, failover, and usage records while Envoy handles the traffic.

SetupDocsCommunityMaturity

Verdict Our Agent Router build succeeded in 273 seconds, but its 12-package test sweep ended 6 passed and 6 failed after 668 seconds, so teams with a green-suite gate should wait or reproduce the failures first. It is a sensible trial for a Kubernetes platform group that already trusts Envoy and needs one policy layer for models and MCP servers.

Full review →

93

skills-hub

qufei1993/skills-hub · ★ 1.7k ▲22/7d · Rust

Skills Hub is a desktop manager for Agent Skills used by Claude Code, Codex, Cursor, and dozens of other coding assistants. It keeps one central copy of each skill, then links or copies it into global and project folders so users can install, tag, update, enable, and remove skills from one interface.

SetupDocsCommunityMaturity

Verdict Our Skills Hub run installed 364 packages, used 261 MB, and passed all 48 tests, but npm audit found 15 known vulnerabilities, 12 of them high severity. It is a good desktop choice for one developer juggling Claude Code, Codex, Cursor, and other local skill folders, provided every imported skill and update is reviewed.

Full review →

94

headcount

cbrock84/headcount · ★ 1.6k · Markdown

Headcount is a Claude Code plugin marketplace that packages 172 operating playbooks as 16 company departments, covering work such as engineering, finance, security, hiring, and marketing. Each skill is a set of instructions Claude loads when a matching request appears. It solves the problem of maintaining one oversized prompt by letting you install only the departments your work needs.

SetupDocsCommunityMaturity

Verdict Our Headcount run installed 35 packages in 25 seconds and built in 8 seconds, but it had no test target for the behavior of its 172 skills. Try it if Claude Code is already part of your work and you want a curated operating method across several departments.

Full review →

95

atlassian-mcp-server

atlassian/atlassian-mcp-server · ★ 1.1k ▲79/7d · JavaScript

Atlassian Rovo MCP Server is Atlassian's cloud-hosted bridge between AI clients and Jira, Confluence, Jira Service Management, Bitbucket Cloud, and Compass. It lets an agent search, read, create, and update work using the user's existing Atlassian permissions; this repository mainly packages connection manifests and reusable skills for the hosted service.

SetupDocsCommunityMaturity

Verdict Our repository run installed only 2 packages and passed 17 of 17 tests, confirming that the open checkout is a small integration package rather than the hosted Atlassian service. Use the official server for supervised Jira and Confluence work when your organization already lives in Atlassian Cloud and admins can control access.

Full review →

96

anolisa

alibaba/anolisa · ★ 615 ▲11/7d · Rust

ANOLISA is a collection of Linux-focused tools for running AI agents from a terminal, trimming tool output, keeping memory, tracing activity, and putting boundaries around execution. It addresses the operational work between an agent framework and the machine it controls.

SetupDocsCommunityMaturity

Verdict Our agent-memory build took 143 seconds and all 692 tests passed, making ANOLISA worth a trial for a Linux team that wants more control around existing agents. The result covers one component in a 12-component repository, so it cannot certify the shell, eBPF tracer, security runtime, or checkpoint daemon.

Full review →

97

browser-use

browser-use/browser-use · ★ 115k ▲580/7d · Python

Browser Use is a Python library and coding-agent skill that lets a language model inspect web pages, click controls, enter text, download files, and return structured results. It is for browser tasks where the steps cannot be captured easily with fixed selectors, and it can run with a local browser or the company's hosted browser service.

SetupDocsCommunityMaturity

Verdict Our Browser Use install occupied 535 MB, reported 19 known vulnerabilities, and never reached tests because pytest rejected a configured argument. It is still a strong experiment for supervised, messy web tasks where fixed scripts break too often, provided the team pins dependencies and tests its own sites.

Full review →

98

codegraph

colbymchenry/codegraph · ★ 71k ▲840/7d · C

CodeGraph builds a local SQLite graph of a codebase, then gives coding agents one MCP tool for retrieving relevant source, call paths, and a change's likely blast radius. It targets the discovery work that agents usually do with repeated searches and file reads, while also offering CLI queries and a browser viewer on the main branch.

SetupDocsCommunityMaturity

Verdict Our CodeGraph checkout installed in 10 seconds and built in 21 seconds, but Vitest failed 852 of 4,618 tests. The local MCP design is worth testing on a copy of a real repository, especially for agents that waste turns rediscovering call paths.

Full review →

99

OpenViking

volcengine/OpenViking · ★ 38k ▲1.1k/7d · Python

OpenViking is a context database for AI agents that stores resources, memories, and skills behind file-like `viking://` paths. It gives an agent one place to browse, search, and retain context, with shorter summaries loaded before full documents.

SetupDocsCommunityMaturity

Verdict Our OpenViking install took 86 seconds and its 13-second build passed, but the test command stopped after 8 seconds because `pytest_asyncio` was missing. It is worth testing when several agent clients need one inspectable context store and your team can own the provider and ingestion machinery.

Full review →

100

CodexPlusPlus

BigPizzaV3/CodexPlusPlus · ★ 31k ▲339/7d · Rust

Codex++ is primarily documented in Chinese, with a full English README available. It is an external launcher and manager for the OpenAI Codex and ChatGPT desktop app, adding provider switching, local protocol conversion, session tools, and interface modifications without patching the official app files.

SetupDocsCommunityMaturity

Verdict Our Codex++ build failed after 212 seconds because its expected frontend `dist` directory was absent, and tests hit the same error after 78 seconds. The release installer may still suit a Codex power user who needs provider switching and session controls, but source contributors should treat the documented frontend build order as mandatory.

Full review →

101

graphiti

getzep/graphiti · ★ 31k ▲123/7d · Python

Graphiti is a Python framework that turns conversations and other changing data into a knowledge graph for AI agents. It keeps the source episode, the extracted entities, and the time window for each fact, which lets an application ask what is true now or what was true earlier.

SetupDocsCommunityMaturity

Verdict Our Graphiti install took 111 seconds and 5,322 MB, then the test suite timed out at 900 seconds after showing errors and failures. Use it when temporal validity, provenance, and graph retrieval justify running a database plus model services.

Full review →

102

dsh-desktop

anywhere-labs/deepseek-harness-desktop · ★ 27k ▲1.3k/7d · TypeScript

DSH Desktop is a Chinese-first desktop wrapper for DeepSeek Harness, with an English README and English release notes available. The repository now redirects to `anywhere-labs/dsh-desktop`; it packages the upstream agent, web interface, host service, and plugin system as Windows and macOS applications with profiles, a tray, terminal access, updates, and a plugin market.

SetupDocsCommunityMaturity

Verdict Our DSH Desktop checkout built in 25 seconds and 770 of 776 tests passed after a 1,352-package install, leaving one lifecycle-evidence failure to investigate. Use the signed app if you already chose DeepSeek Harness and want a native Windows or macOS shell.

Full review →

103

llamafile

mozilla-ai/llamafile · ★ 26k ▲24/7d · C++

llamafile packs a large language model and its runtime into a single executable file that runs on Windows, macOS, Linux, and BSD without installation. It also includes whisperfile, a single-file speech-to-text tool built on whisper.cpp. The goal is to make open LLMs easy to distribute and run locally.

SetupDocsCommunityMaturity

Verdict Our test run ended with 7 passed and 2 errors from missing wget and appium modules, so llamafile is not a turnkey development checkout despite its one-file runtime pitch. For end users grabbing a prebuilt llamafile, it is still a compelling way to run an LLM locally without installing anything.

Full review →

104

distilly

titanwings/distilly · ★ 25k ▲174/7d · TypeScript

Distilly turns selected writing, transcripts, public pages, and working material into a versioned profile of how a person thinks and communicates. An agent can load that profile for one run or install an approved copy as a lasting local Skill, with the source material and history kept on the user's machine.

SetupDocsCommunityMaturity

Verdict Our Distilly run installed 252 packages in 27 seconds and built in 19 seconds, but tests timed out at 900 seconds after the log had shown a state-store failure. The evidence model and explicit review flow are thoughtful, yet this remains source-installed preview software tied to recorded host versions.

Full review →

105

NemoClaw

NVIDIA/NemoClaw · ★ 22k ▲35/7d · TypeScript

NVIDIA NemoClaw is a reference stack for running OpenClaw, Hermes, or LangChain Deep Agents Code inside NVIDIA OpenShell sandboxes. Its CLI combines onboarding, managed inference, network policy, integrations, snapshots, and lifecycle operations so an agent gets bounded host and network access.

SetupDocsCommunityMaturity

Verdict Our NemoClaw install pulled 387 packages in 47 seconds, but the 36-second test step failed because TypeScript could not resolve `json5` and `tar`. Evaluate it only when a supported NVIDIA or WSL host and one of its 3 named agents match your deployment.

Full review →

106

omlx

jundot/omlx · ★ 22k ▲247/7d · Python

oMLX is a local inference server for running language, vision, embedding, and reranking models on Apple Silicon Macs. It supplies OpenAI- and Anthropic-compatible APIs, continuous batching, RAM and SSD prompt caching, model management, and a native menu-bar application.

SetupDocsCommunityMaturity

Verdict Our oMLX environment occupied 5,842 MB and built in 8 seconds, but tests stopped after 7 seconds because `libmlx.so` could not be loaded on Debian. Use the packaged Mac app when an Apple Silicon workstation should serve several local models to coding tools and compatible clients.

Full review →

107

adk-python

google/adk-python · ★ 22k ▲52/7d · Python

Google's Agent Development Kit is a Python framework for defining agents, tools, sessions, evaluations, and multi-agent workflows in code. It is built around Gemini but can work with other model providers, and it gives teams a path from a local command-line agent to Cloud Run or Vertex AI Agent Engine.

SetupDocsCommunityMaturity

Verdict Our ADK install pulled 331 packages and used 1,696 MB, while pytest stopped at 200 collection or setup errors, so this is a serious framework that deserves a trial before adoption. Choose it when Gemini, explicit workflow graphs, evaluation, and Google Cloud deployment belong in the same Python stack.

Full review →

108

jcode

1jehuang/jcode · ★ 20k ▲268/7d · Rust

jcode is a Rust terminal coding agent with persistent sessions, semantic memory, multiple model providers, browser control, and coordinated agent swarms. It is meant to replace a pile of separate coding-agent utilities with one client that can run interactively, headlessly, or through a persistent server.

SetupDocsCommunityMaturity

Verdict Our jcode run built in 660 seconds, then 5 of 231 tests failed, so this is an ambitious coding environment that still asks adopters to tolerate breakage at the edges. Try it if persistent multi-session work, provider choice, and coordinated agents solve problems you already have.

Full review →

109

llm_wiki

nashsu/llm_wiki · ★ 20k ▲930/7d · TypeScript

LLM Wiki is a desktop app that turns imported documents into a linked, searchable wiki maintained with language models. It keeps the source files, generated pages, citations, graph, and chat history on your machine, while letting you choose local or hosted model providers.

SetupDocsCommunityMaturity

Verdict Our llm_wiki run installed 862 packages and 621 MB, then its test command exited 1 despite a 1,876-pass summary, so this is an ambitious desktop knowledge system that still needs careful evaluation. Use it if generated Markdown, source traceability, and local agent access solve a real research problem for you.

Full review →

110

a2ui

a2ui-project/a2ui · ★ 16k ▲52/7d · TypeScript

A2UI is a protocol and set of renderer libraries for letting an AI agent describe an interactive interface as JSON. The client maps that description onto a catalog of approved components, which gives remote agents forms, cards, and other UI without executing code they generated.

SetupDocsCommunityMaturity

Verdict Our A2UI install pulled 2,547 packages, occupied 3,045 MB, and took 154 seconds, while our harness found no build or test target to run. Adopt v0.9.1 for an evaluated agent-UI project when cross-client portability and a client-owned component catalog justify that weight.

Full review →

111

aisuite

andrewyng/aisuite · ★ 16k ▲17/7d · Python

aisuite is a Python library that gives several hosted and local language-model providers one OpenAI-shaped chat interface. It also has an agent runner, tool policies, persistent state, streaming, and MCP connections for developers who want one application to switch among providers.

SetupDocsCommunityMaturity

Verdict Our aisuite run installed 136 packages in 78 seconds, but 43 tests failed with repeated messages that async test functions were not natively supported. Use it for a Python application that genuinely needs provider switching or MCP tools and can maintain provider-level contract tests.

Full review →

112

claude-obsidian

AgriciDaniel/claude-obsidian · ★ 15k ▲236/7d · Python

claude-obsidian is a local knowledge system that uses Claude Code or another Agent Skills host to turn source material into linked, cited Obsidian notes. It keeps Markdown, captured sources, provenance records, and transaction data in a vault the user owns.

SetupDocsCommunityMaturity

Verdict Our claude-obsidian run passed 458 of 463 tests but failed 5 after 579 seconds, so its careful write design did not produce a clean baseline at commit 1c1bc49. It is a serious choice for Claude Code users who will review every mutation plan and value plain-file provenance.

Full review →

113

geo-seo-claude

zubair-trabzada/geo-seo-claude · ★ 11k ▲109/7d · Python

GEO-SEO Claude Code is a Claude Code skill that audits websites for visibility in AI answers while also checking familiar SEO basics. It turns page fetching, crawler rules, schema checks, content scoring, and report generation into slash commands for people who already work inside Claude Code.

SetupDocsCommunityMaturity

Verdict Our geo-seo-claude run installed 50 packages in 47 seconds and passed all 14 tests, making it cheap for a Claude Code user to try. Use it as a structured first pass for English-language sites, then verify the fetched page, headers, schema, and business claims before sending a report.

Full review →

114

omnigent

omnigent-ai/omnigent · ★ 10k ▲156/7d · Python

Omnigent is an alpha-stage control layer for running Claude Code, Codex, Cursor, OpenCode, Hermes, Pi, and custom agents through shared sessions. It gives teams one place for terminals, web and mobile access, policies, cloud sandboxes, and multi-agent delegation.

SetupDocsCommunityMaturity

Verdict Our Omnigent checkout pulled 1,712 packages and used 2,086 MB, then offered no root build or test target, so adopting it means evaluating a large alpha platform beyond the quick-start path. Try it if several coding agents, shared sessions, and enforceable policies have already become one operating problem for your team.

Full review →

115

baml

BoundaryML/baml · ★ 9.2k ▲15/7d · Rust

BAML is a programming language and toolchain for defining typed LLM functions, prompts, provider settings, tests, and generated application clients. It is meant to replace hand-written provider calls and JSON cleanup with one schema that can generate clients for languages including Python, TypeScript, Go, Java, C#, Ruby, and Rust.

SetupDocsCommunityMaturity

Verdict Our BAML checkout installed 2,528 packages and consumed 3,740 MB, then both build and test failed at the Python language-client task, so source contributors need more patience than the user quickstart suggests. BAML is worth a trial when structured LLM calls have spread across providers and languages, because one typed definition can replace duplicated prompt and parsing code.

Full review →

116

slime

THUDM/slime · ★ 8.5k ▲65/7d · Python

slime is a framework for improving large language models with reinforcement learning after initial training. It connects Megatron training, SGLang rollouts, rewards, checkpoints, and custom agent environments in one distributed loop for teams that would otherwise assemble those systems themselves.

SetupDocsCommunityMaturity

Verdict Our slime run installed 191 packages and built successfully, but pytest finished with 26 failures and 29 errors after 392 seconds. Adopt it only when Megatron and SGLang are deliberate choices and your team can investigate distributed training failures at model, kernel, and cluster level.

Full review →

117

munder-difflin

chaitanyagiri/munder-difflin · ★ 7.4k ▲707/7d · JavaScript

Munder Difflin is a local desktop control room for running several terminal coding agents, giving each one a mailbox, shared memory, worktree, task board, and pixel-art desk. It solves the coordination problem that appears when Claude Code, Codex, Copilot, and other command-line agents work in parallel and someone still needs to route tasks, watch costs, and review their work.

SetupDocsCommunityMaturity

Verdict Our Munder Difflin build passed, but its 781-package install used 1,062 MB and npm audit found 28 vulnerabilities, including 1 critical and 23 high. Use it as a supervised control room on backed-up repositories if mailboxes, worktrees, visible terminals, and budgets solve a real coordination problem.

Full review →

118

code-graph-rag

vitali87/code-graph-rag · ★ 5.1k ▲19/7d · Python

Code-Graph-RAG turns a mixed-language repository into a graph of files, symbols, calls, imports, and data flow, then lets people or coding agents query and edit it. It uses Tree-sitter and language-specific analysis, stores the graph in Memgraph, and can expose its tools to Claude Code and other MCP clients.

SetupDocsCommunityMaturity

Verdict Our Code-Graph-RAG run installed 150 packages and built in 10 seconds, but pytest stopped at 200 Docker connection errors after only 24 passes, so Docker is a real prerequisite rather than an optional extra. Try it when cross-language relationships would change how your team searches a monorepo and you can operate the graph services.

Full review →

119

aidlc-workflows

awslabs/aidlc-workflows · ★ 4.7k ▲103/7d · TypeScript

AI-DLC Workflows is an AWS Labs framework that makes coding agents follow a recorded software-development process instead of jumping straight from a prompt to code. One shared TypeScript engine supplies staged planning, specialist agents, human approvals, audit records, and resumable state to 7 supported coding assistants.

SetupDocsCommunityMaturity

Verdict Our AI-DLC checkout installed 129 packages in 124 seconds, but the lab found no build or test target to execute, so it produced no passing suite result. Trial it when auditability and human gates are requirements, and budget time for Bun, harness-specific configuration, repository files, and provider credentials.

Full review →

120

dictionary-of-ai-coding

mattpocock/dictionary-of-ai-coding · ★ 4.7k · TypeScript

Dictionary of AI Coding is an English-language glossary for the terms developers meet while working with coding agents. It explains models, tokens, context, tools, failure modes, handoffs, memory, review, and agent-friendly development in plain language with usage examples.

SetupDocsCommunityMaturity

Verdict Our 6-second install pulled 42 packages and 49 MB, while the repository offered no build or test target, confirming that this is a reading resource with a small generator rather than a software tool. Use it when a team needs sharper language for coding-agent work, especially around context, handoffs, and review.

Full review →

121

OpenResearch

alphaXiv/OpenResearch · ★ 4.4k ▲3.1k/7d · Rust

OpenResearch is a local workspace for sending coding agents through literature review, hypothesis development, experiments, and research writing. It solves the messy coordination problem around parallel agent sessions by keeping their code, evidence, runs, and outputs together on your machine.

SetupDocsCommunityMaturity

Verdict Our build took 125 seconds, and all 1,370 tests passed. OpenResearch is worth trying if agent-led research in code repositories has outgrown a pile of terminals, branches, and loose notes.

Full review →

122

freecad-mcp

neka-nat/freecad-mcp · ★ 2.4k ▲203/7d · Python

FreeCAD MCP lets an AI client such as Claude Desktop create and inspect models inside the FreeCAD desktop app. A Python server translates the client's requests into FreeCAD operations, so the model can add parts, edit properties, take views, and run scripts without the user clicking through every command.

SetupDocsCommunityMaturity

Verdict Our 69 MB install passed all 41 tests, but using FreeCAD MCP still means running a desktop GUI that can execute agent-supplied Python with full FreeCAD privileges. It is a good trial for an experienced FreeCAD user who wants faster conversational iteration and will inspect the resulting geometry.

Full review →

123

opencode-mem

tickernelz/opencode-mem · ★ 1.7k ▲13/7d · TypeScript

OpenCode Memory is a plugin that gives the OpenCode coding agent searchable memory across sessions. It stores project decisions, prior fixes, prompts, and a user-preference profile in local libSQL shards, then injects relevant items into later chats.

SetupDocsCommunityMaturity

Verdict Our run installed 128 packages and 510 MB, then the build failed and 9 tests failed, so this commit is a trial candidate rather than a default recommendation. The migration safeguards and local libSQL design are thoughtful, and 435 tests did pass.

Full review →

124

reversa

sandeco/reversa · ★ 1.6k · JavaScript

Reversa installs a collection of coding-agent skills into an existing project, then guides an agent through extracting business rules, architecture, data relationships, and gaps from the code. It is meant for teams that inherited a system without dependable specifications and want traceable working documents before they change or rebuild it.

SetupDocsCommunityMaturity

Verdict Our 80-second Reversa install added 61 packages and passed an audit with 0 known vulnerabilities, but the measured commit exposed no build or test target for us to run. Use it when a neglected codebase needs a structured first map and every generated claim will receive human review.

Full review →

125

sf-skills

forcedotcom/sf-skills · ★ 1.0k ▲23/7d · Python

Salesforce Skills Library is a large collection of Agent Skills for building and operating Salesforce applications. Its folders give compatible coding agents instructions, scripts, references, and assets for Apex, Flow, SOQL, Lightning, Agentforce, metadata, permissions, integrations, and other Salesforce work.

SetupDocsCommunityMaturity

Verdict Our sf-skills checkout installed 248 packages in 20 seconds, but npm offered no build or test target for its 7,244 files. Salesforce teams should browse and pin the specific skills they need rather than absorb the whole fast-changing catalog without review.

Full review →

126

Automodel

NVIDIA-NeMo/Automodel · ★ 957 ▲29/7d · Python

NeMo AutoModel is NVIDIA's PyTorch training library for pretraining and fine-tuning language, vision-language, diffusion, and retrieval models from Hugging Face checkpoints. YAML recipes separate model and training choices from distributed layouts, so the same code can move between one GPU and multi-node jobs.

SetupDocsCommunityMaturity

Verdict Our AutoModel environment occupied 5,703 MB and 1,980 tests passed, but 153 failed and 47 hit collection/setup errors, so adoption requires a pinned recipe-specific environment rather than blind trust in the default checkout. Use it when NVIDIA GPU scale, Hugging Face checkpoint compatibility, and advanced parallel layouts justify dedicated ML platform work.

Full review →

127

funes

huggingface/funes · ★ 427 ▲49/7d · Rust

Funes turns past Claude Code, Codex, pi, and Hermes sessions into a searchable local memory that an AI coding agent can query later. It solves the familiar problem of decisions, dead ends, and useful findings disappearing when a session closes, while optionally sharing that history through a Hugging Face dataset.

SetupDocsCommunityMaturity

Verdict Our Funes run built in 773 seconds and passed all 554 tests in another 283 seconds, a strong result for a young Rust project with a heavy compile. Use it if lost coding-agent context is costing real time and you can keep memory local until a deliberate sharing review.

Full review →

128

LlamaFactory

hiyouga/LlamaFactory · ★ 75k ▲104/7d · Python

LlamaFactory is a Python toolkit for adapting more than 100 large language and multimodal model families to your own data. It brings command-line tools, a browser interface, multiple training methods, experiment tracking, and serving options into one project, reducing the amount of separate training plumbing a team must assemble.

SetupDocsCommunityMaturity

Verdict Our build completed in 35 seconds, but tests ended with 3 collection/setup errors after 318 tests passed. LlamaFactory is a compelling shortlist choice when breadth matters: it combines many model families, training methods, interfaces, and inference paths in one active project.

Full review →

129

nanoGPT

karpathy/nanoGPT · ★ 63k ▲153/7d · Python

nanoGPT is a small Python codebase for training, fine-tuning, and sampling GPT-style language models without a large framework around the model loop. It solves the problem of learning from or changing a readable implementation, but its maintainer now labels the project deprecated and directs new users to nanochat.

SetupDocsCommunityMaturity

Verdict Our nanoGPT run installed 35 packages in 32 seconds and built in 21 seconds, but it had no test target, so the clean setup does not prove the training code still works across supported hardware. Use it to read, modify, or reproduce its compact GPT implementation.

Full review →

130

minimind

jingyaogong/minimind · ★ 61k ▲626/7d · Python

MiniMind is a Chinese-first tutorial repository for training a small language model from scratch; a full English README is provided beside the main Chinese guide. It exposes the model, tokenizer, data loaders, alignment methods, local inference, and API code so learners can follow the whole path instead of calling a high-level training wrapper.

SetupDocsCommunityMaturity

Verdict Our MiniMind install occupied 5,412 MB and pip-audit reported 78 known vulnerabilities, so use it in an isolated learning environment rather than as a production dependency. Its 4,669 lines of source give learners an unusually broad view of language-model training without hiding every algorithm behind a framework.

Full review →

131

supermemory

supermemoryai/supermemory · ★ 30k ▲109/7d · TypeScript

Supermemory stores facts and documents for AI applications, then retrieves relevant history and maintains a changing user profile across conversations. Developers can use its hosted API, connect AI clients through MCP, or run a smaller single-machine server with their own language model.

SetupDocsCommunityMaturity

Verdict Our bun install pulled 3,015 packages and used 3,212 MB, then the 98-second build failed because its postbuild required a Sentry auth token. Supermemory is a sensible trial for an agent product that needs profiles plus document retrieval, especially if the hosted connectors remove work you already planned to do.

Full review →

132

skills

openai/skills · ★ 27k ▲451/7d · Python

OpenAI Skills is an English-language catalog of instruction folders, scripts, and reference material that Codex can load for repeatable tasks. Its README now marks the repository deprecated and directs new authors to OpenAI Plugins and the current Codex skill documentation.

SetupDocsCommunityMaturity

Verdict OpenAI has deprecated this 44-package catalog, and the README now sends new Codex extension work to `openai/plugins`. Read its skill packages for examples or to maintain an old setup, but do not choose it as the home of a new distributable skill.

Full review →

133

gastown

gastownhall/gastown · ★ 18k ▲72/7d · Go

Gas Town is a workspace manager for coordinating many coding agents across one or more Git repositories. It stores tasks, identities, handoffs, and recovery state outside an agent's chat so work can survive restarted sessions.

SetupDocsCommunityMaturity

Verdict Our Gas Town run installed 651 packages and built in 108 seconds, but 27 of 75 Go test packages failed. Its persistent work model is a credible answer to 20-agent coordination, yet the setup and red suite make it a poor default for small teams.

Full review →

134

cc-connect

chenhg5/cc-connect · ★ 16k ▲61/7d · Go

CC-Connect is a Go bridge that lets you operate local AI coding agents from chat services instead of staying at the development machine. It connects tools such as Claude Code, Cursor, Gemini CLI, and Kimi CLI to services including Slack, Discord, Telegram, Feishu/Lark, DingTalk, and WeCom.

SetupDocsCommunityMaturity

Verdict Our build succeeded in 11 seconds, but 7 of 93 tests failed. CC-Connect is a compelling shortcut if you genuinely need one local coding agent across several chat networks, and its active repository makes it worth a trial.

Full review →

135

DesktopCommanderMCP

wonderwhy-er/DesktopCommanderMCP · ★ 9.6k ▲65/7d · TypeScript

Desktop Commander is an MCP server that lets an AI client read and edit local files, search folders, run terminal commands, manage processes, and work with spreadsheets, PDFs, and Word documents. It turns Claude Desktop, Claude Code, Codex, and other MCP clients into operators of the computer where the server runs.

SetupDocsCommunityMaturity

Verdict Our Desktop Commander install ran for 900 seconds without completing, so we could not build or test the 301-file checkout. The tool is useful when a trusted MCP client genuinely needs terminal, process, filesystem, and office-document access, but host mode grants more authority than its directory and command settings can contain.

Full review →

136

PraisonAI

MervinPraison/PraisonAI · ★ 9.1k ▲17/7d · Python

PraisonAI is a Python and JavaScript framework for building agents that call language models, tools, MCP servers, memory stores, and other agents. It packages single-agent prompts, multi-agent graphs, approval hooks, execution limits, dashboards, and remote runtimes under one project.

SetupDocsCommunityMaturity

Verdict Our PraisonAI npm audit found 48 known vulnerabilities, including 3 critical and 29 high, while the measured package exposed no build or test target. Its feature range is useful for experiments, but that dependency result and the open safety-gate report rule out trusting autonomous production tools without a separate security review.

Full review →

137

gentle-ai

Gentleman-Programming/gentle-ai · ★ 6.9k ▲241/7d · Go

Gentle AI configures AI coding tools already installed on a developer's machine. It adds shared workflows, persistent memory, optional spec-driven work, skills, MCP connections, model routing, personas, and an opt-in review system across clients such as Claude Code, OpenCode, Cursor, Codex, and Gemini CLI. It is not an AI agent runtime.

SetupDocsCommunityMaturity

Verdict Our harness installed 0 Node packages and failed in 6 seconds because the root npm test is a placeholder, so that result does not validate Gentle AI's Go implementation. Adopt v2.4.0 only if cross-agent configuration and opt-in evidence review solve a problem your existing repository policy does not.

Full review →

138

Kun

KunAgent/Kun · ★ 6.3k ▲4/7d · TypeScript

Kun is a Chinese-first, local AI-agent workbench for coding, design, writing, document analysis, and automation through a shared desktop and terminal runtime. A full English README and English guides exist, so English-speaking users can evaluate and operate it without translating the main Chinese page.

SetupDocsCommunityMaturity

Verdict Our Kun install consumed 2,374 MB, and its wider suite finished with 85 failed tests plus 114 known vulnerabilities, so v0.3.6 is not a low-risk default for sensitive work. Noncommercial users may still find the shared GUI and TUI runtime unusually coherent if they can isolate tools, review every permission, and accept a fast-moving release.

Full review →

139

awesome-deepseek-agent

deepseek-ai/awesome-deepseek-agent · ★ 6.1k ▲27/7d

Awesome DeepSeek Agent is a bilingual collection of setup guides for using DeepSeek models inside coding assistants and general agent tools. It solves a fiddly configuration problem by putting working installation, API-key, model, and first-run instructions for 22 tools in one place.

SetupDocsCommunityMaturity

Verdict Awesome DeepSeek Agent is a useful first stop if your exact client appears in its table, especially because the guides are short and bilingual. It is a reference shelf, not a product, and its large unmerged contribution queue makes freshness uneven.

Full review →

140

ouroboros

Q00/ouroboros · ★ 6.0k · Python

Ouroboros is a local Python workflow layer for AI coding agents. It interviews you about a task, writes the answers into a fixed specification called a Seed, hands work to an existing coding runtime, records what happened, and checks the result through staged evaluation. English is the primary README, with Korean and Simplified Chinese versions also available.

SetupDocsCommunityMaturity

Verdict Our Ouroboros run installed 92 packages in 28 seconds and built in 6 seconds, but tests stopped after 1 second because `pytest_asyncio` was missing. Use it when vague requirements and weak evidence are costing more time than the coding itself, and you want the agent forced through a recorded contract.

Full review →

141

open-knowledge

inkeep/open-knowledge · ★ 4.2k ▲48/7d · TypeScript

OpenKnowledge is a desktop and local web editor that makes Markdown and MDX folders feel closer to Notion while keeping files in an ordinary project directory. It adds graph navigation, Git/GitHub sync, search, and side-by-side work with Claude, Codex, and other agents through MCP, CLI, and skills.

SetupDocsCommunityMaturity

Verdict Our OpenKnowledge run installed 1,934 packages and used 1,829 MB, then both build and test commands failed after 7 seconds in the native-config path. Trial the v0.67.2 desktop installer if you want a visual Markdown workspace with Claude Code and Codex access, since the lab failure covers the source checkout rather than the published installers.

Full review →

142

lean-ctx

yvgude/lean-ctx · ★ 3.8k ▲30/7d · Rust

LeanCTX is a local Rust layer between AI coding agents and the code, shell output, and model requests they consume. Its MCP server, shell hooks, session store, and optional provider proxy try to send less repeated material while preserving a route back to the original content. It also records context budgets, token use, and accepted outcomes.

SetupDocsCommunityMaturity

Verdict Our LeanCTX install took 59 seconds and the 537-second build passed, but tests exited 101 after another 196 seconds, so source adopters do not yet have a clean result from our sandbox. It is worth a controlled trial for heavy Claude Code or Codex use in a large repository, especially when repeated reads and command output are the problem.

Full review →

143

claudish-to-english

gvzdv/claudish-to-english · ★ 2.6k ▲51/7d · Shell

claudish-to-english is a Claude Code plugin that rewrites assistant messages in plain English before displaying them. It leaves Claude's original response and saved transcript untouched, while using Ollama locally by default or an Anthropic or OpenAI-compatible API to produce a more readable version.

SetupDocsCommunityMaturity

Verdict Use it if Claude Code's wording regularly slows you down and you already run Ollama or have a suitable API account. The display-only, fail-open architecture is thoughtfully conservative, and the documentation is unusually candid about prerequisites.

Full review →

144

oh-my-hermes

rlaope/oh-my-hermes · ★ 2.5k ▲867/7d · Python

oh-my-hermes, or OMH, is an operating layer for NousResearch's Hermes Agent. It installs skills, workflow routing, model preferences, local memory contracts, a terminal display, and guarded handoffs to coding tools such as Claude Code and Codex while trying to distinguish prepared work from observed results.

SetupDocsCommunityMaturity

Verdict Our oh-my-hermes test run reached 34% and timed out after 900 seconds with failure markers already in the log. Existing Hermes users may still value its explicit evidence states and guarded coding handoffs, but v2.0.0 is too large and too lightly explained to install on trust.

Full review →

145

webcodex

yyjeqhc/webcodex · ★ 964 ▲340/7d · Rust

WebCodex is an open-source Rust tool that gives AI assistants like ChatGPT and Claude a local, permission-bounded development environment on your own machine. It solves the problem of moving a repository into a hosted AI workspace by keeping code, Git, tests, and developer tools where they already live, connected over MCP or HTTPS.

SetupDocsCommunityMaturity

Verdict WebCodex v0.4.0 built cleanly in 246 seconds and passed 4,958 of 4,962 tests in our sandbox, but the four failing Git tests and a 623-star footprint make it a cautious try for local AI agent work, not a safe default for production. Try the temporary `share` command first; only invest in the Desktop or Server + Runner setup if you need durable, reviewable long-running agent sessions on your own machines..

Full review →

146

ambient-context

dragthelake/ambient-context · ★ 171 ▲4/7d · Rust

Ambient Context is a macOS menu bar app that records text from your focused window into local Markdown files, then lets an AI assistant use that record as work memory. It is for people tired of re-explaining the day's tabs, messages, decisions, and unfinished tasks to a coding agent.

SetupDocsCommunityMaturity

Verdict Our Ambient Context run installed 79 packages in 5 seconds and built in 6 seconds, but commit 6a353ab supplied no test target, so we could not verify behavior with an automated suite. Use it if you want an agent-readable diary on an Apple Silicon Mac and can protect the plaintext folder it creates.

Full review →

147

deepseek-harness

deepseek-ai/deepseek-harness · ★ 227k ▲9.2k/7d · TypeScript

DeepSeek Harness is an open source coding-agent runtime with a local web interface, headless mode, model adapters, tools, sessions, approvals, sandboxing, and MCP client support. Its defining idea is that nearly every part of the agent loop is a replaceable Cordis plugin, so developers can change behavior through composition instead of forking a fixed core.

SetupDocsCommunityMaturity

Verdict Our DeepSeek Harness run passed 14,583 of 14,707 tests, with 10 failures and 114 skips, so the large suite was close but not clean. Use it to study or build a replaceable agent runtime when the Cordis plugin model is the point.

Full review →

148

autogen

microsoft/autogen · ★ 61k ▲68/7d · Python

AutoGen is a Python and .NET framework for building agents that exchange messages, call tools, and divide work across a team. Microsoft has put it in maintenance mode and directs new projects to Microsoft Agent Framework, so AutoGen now makes most sense for existing deployments and migration work.

SetupDocsCommunityMaturity

Verdict Our AutoGen build passed in 10 seconds, but its test run produced 96 collection or setup errors and no passing tests, while Microsoft now labels the project maintenance-only. Keep it for an existing application that cannot move yet, with targeted tests around the packages you use.

Full review →

149

continue

continuedev/continue · ★ 36k ▲62/7d · TypeScript

Continue is an Apache-licensed coding agent available as a command-line tool, a VS Code extension, and a JetBrains plugin. It lets developers connect a coding workflow to language models, but its maintainers now describe the repository as read-only and version 2.0.0 as the final release.

SetupDocsCommunityMaturity

Verdict Our Continue install took 16 seconds and reported 0 known npm vulnerabilities, but the maintainers now call 2.0.0 the final release and the repository read-only. Existing users have a usable endpoint to pin, while new users would be adopting 3,001 files of agent code without normal upstream maintenance.

Full review →

150

airllm

lyogavin/airllm · ★ 34k ▲254/7d · Jupyter Notebook

AirLLM runs language models that are too large to fit in GPU memory by loading their weights a layer, or one routed expert, at a time. It trades inference speed and heavy disk use for the chance to run models such as 70B-class systems on consumer hardware.

SetupDocsCommunityMaturity

Verdict Our AirLLM install took 120 seconds and its build took 16 seconds, but the repo gave us no test target, so it is best treated as an experiment for running oversized models rather than a proven serving layer. Use it when GPU memory is the hard limit and slow, storage-heavy inference is acceptable.

Full review →

151

RAG_Techniques

NirDiamant/RAG_Techniques · ★ 30k · Jupyter Notebook

RAG_Techniques is an English-language collection of 42 Jupyter notebooks that teach ways to retrieve better context for language-model answers. It helps you compare methods such as chunking, reranking, Graph RAG, and evaluation before choosing what to build. It is a teaching library rather than an installable RAG product.

SetupDocsCommunityMaturity

Verdict Our 17-second test attempt on RAG_Techniques stopped before collection because `langchain_openai` was missing, so the repository is easier to read than to verify as one working unit. Use its 42 notebooks to select and understand a technique, then rebuild that technique with pinned dependencies and tests for your own data.

Full review →

152

patent-disclosure-skill

handsomestWei/patent-disclosure-skill · ★ 9.7k ▲449/7d · Python

Patent Disclosure Skill is a Chinese-language Claude Code and Cursor skill for drafting Chinese patent disclosure documents, reading published patents, watching policy changes, and assisting with office-action replies. The primary README, installation guide, and workflow instructions are in Chinese; there is a short English description in the skill metadata, but no full English documentation path.

SetupDocsCommunityMaturity

Verdict Our run passed 179 of 182 collected tests, but 2 tests failed and 1 collection or setup error remained, so Patent Disclosure Skill should be treated as a review-heavy drafting aid. Chinese-speaking patent teams may save clerical effort because it joins source scanning, CNIPA search, diagrams, Word output, and revision files in one workflow.

Full review →

153

dsh-routing-suite

yjh051108/dsh-routing-suite · ★ 7.2k ▲22/7d · JavaScript

dsh-routing-suite is a Chinese-first extension suite for DeepSeek Harness, with a complete English README available. It combines a runtime plugin injector with presets that classify a task and change the agent's reasoning persona, tool exposure, and reminders during a session.

SetupDocsCommunityMaturity

Verdict Our dsh-routing-suite build stopped after 6 seconds because it could not find the required DSH checkout, even though its 37-package npm install and audit completed cleanly. Use it only inside an existing DeepSeek Harness setup where you can inspect request headers, tool catalogs, and API spend after every update.

Full review →

154

Switchyard

NVIDIA-NeMo/Switchyard · ★ 3.1k ▲235/7d · Rust

Switchyard is a Rust proxy and library that routes language-model requests across providers while translating OpenAI and Anthropic API formats. It lets an application keep one client protocol while choosing backends through random splits, classifiers, conversation signals, or custom Rust code.

SetupDocsCommunityMaturity

Verdict Our Switchyard run built in 10 seconds, but pytest ended with 86 passes, 10 failures, and 2 collection/setup errors, so commit 053a61e is research-grade rather than a clean production dependency. Try libsy when model routing itself is the work and your team can follow rapid API changes.

Full review →

155

chat-on-steroids

totec448-spec/chat-on-steroids · ★ 2.8k ▲1.1k/7d · TypeScript

Chat On Steroids is an Electron desktop app and local MCP server that lets ChatGPT read files, patch code, run shell commands, retain transcripts, and coordinate worker chats on your computer. It closes the gap between a browser conversation and local development tools, with approved folders and per-capability switches defining what ChatGPT may do.

SetupDocsCommunityMaturity

Verdict Our 60-second install and 14-second build worked, but 1 of 3,184 tests failed, so Chat On Steroids 2.0.7 is a credible beta for developers who specifically want browser ChatGPT to operate a local workspace. Its permission switches and local records are thoughtful, yet the unsigned apps, extension pairing, workspace limits, and browser automation make it a poor default for managed or unattended systems.

Full review →

156

book-to-skill

Leutenegger/book-to-skill · ★ 1.2k · Python

book-to-skill converts technical books and folders of documentation into structured agent skills with an index, chapter notes, a glossary, patterns, and a cheat sheet. It is meant to solve a familiar retrieval problem: useful material is sitting in a file, but finding and applying the relevant section while working takes too much effort.

SetupDocsCommunityMaturity

Verdict Use book-to-skill as an interesting local workflow, not yet as trusted knowledge infrastructure. Its output structure is thoughtful and its format coverage is unusually broad, but extraction quality and generated notes still need human checking.

Full review →

157

northcinder

cinderline/northcinder · ★ 1.2k ▲1/7d · JavaScript

NorthCinder is a local MCP server that helps an AI app compare products, keep source and rejection evidence, and ask for approval before checkout. It is designed as buyer-controlled shopping infrastructure rather than a hosted marketplace, and the repository owner says it operates no NorthCinder cloud service.

SetupDocsCommunityMaturity

Verdict Our NorthCinder run installed 409 packages in 16 seconds and its protocol suite passed 183 tests, but the overall build and test commands still exited 1 because the site path failed. The product's approval receipts, local audit trail, and visible store gaps are sensible safeguards for an experimental shopping agent.

Full review →

158

OpenContext

0xranx/OpenContext · ★ 1.2k ▲100/7d · JavaScript

OpenContext keeps project notes and decisions in a local library so coding assistants can reuse them across sessions and repositories. Its command-line tool, desktop interface, and Model Context Protocol server let Cursor, Claude Code, and Codex find or update that material without replacing the coding agent you already use.

SetupDocsCommunityMaturity

Verdict Our OpenContext install pulled 1,041 packages and 821 MB, while npm audit reported 37 known vulnerabilities including 1 critical, so the dependency tree needs review before this belongs on a sensitive workstation. The local Markdown model and 10 MCP tools are a sensible fit for one developer who wants portable context across coding agents.

Full review →

159

agentic-api

vllm-project/agentic-api · ★ 278 ▲23/7d · Rust

Agentic API is a Rust gateway that adds conversation state, tool loops, and OpenAI Responses or Anthropic Messages protocol handling in front of vLLM. It lets Codex, Claude Code, SDKs, and custom clients use self-hosted models without making every client manage history and server-side tools itself.

SetupDocsCommunityMaturity

Verdict Our Agentic API run installed 36 Python packages and built in 6 seconds, but pytest ended with 2 failures and 1 collection or setup error after 109 tests passed. It is worth a controlled trial for teams already running vLLM and willing to treat the gateway as security-sensitive infrastructure.

Full review →

160

9router

decolua/9router · ★ 29k ▲575/7d · JavaScript

9Router is a self-hosted traffic switchboard for AI coding clients such as Claude Code, Codex, and Cursor. It translates client formats, tracks provider quotas, compresses tool output, and can fall back across subscriptions, paid APIs, free tiers, or local models when a route stops working.

SetupDocsCommunityMaturity

Verdict Our 9Router run installed 871 packages in 99 seconds, occupied 680 MB, and built in 80 seconds, while npm audit found 2 moderate vulnerabilities and no test target ran. That result supports a local trial for a technical user juggling several coding subscriptions or free tiers.

Full review →

161

codex-with-chatgpt

XiaoDuoYa/codex-with-chatgpt · ★ 4.8k ▲757/7d · TypeScript

Codex with ChatGPT is an unofficial TypeScript bridge that lets the ChatGPT web app read selected files and Git state while Codex keeps control of edits and commands. It targets people who want ChatGPT to plan and review coding work without sending a whole repository or using an API key. The README opens in Chinese and English, then provides detailed English instructions and links a full Simplified Chinese guide.

SetupDocsCommunityMaturity

Verdict Our sandbox install failed after 7 seconds on Node.js v20.20.2, so Codex with ChatGPT is not ready to become the default planning layer for a working Codex setup. Try it in a disposable workspace if you accept remote read access and occasional connector repair.

Full review →

162

hyperresearch

jordan-gibbs/hyperresearch · ★ 3.4k ▲992/7d · Python

Hyperresearch is a Python package that turns Claude Code into a multi-stage research runner and keeps gathered sources in a reusable Markdown vault. It searches, drafts, criticizes, checks citations, and resumes interrupted work instead of treating each report as a disposable chat.

SetupDocsCommunityMaturity

Verdict Our Hyperresearch suite was still at 30% when the 900-second cap stopped it, so this is for Claude Code users who value a deep audit trail more than a fast feedback loop. The Markdown vault and explicit citation checks are useful ideas, but 848 MB of installed packages, 1 known vulnerability, and two open runtime defects demand a trial on your own workflow.

Full review →

163

cve-mcp-server

mukul975/cve-mcp-server · ★ 1.6k ▲24/7d · Python

CVE MCP Server gives an MCP-compatible assistant tools for looking up vulnerabilities, exploitation signals, threat data, and package risk. Its main job is to collect evidence from several security services so an analyst can triage a CVE without opening each service by hand.

SetupDocsCommunityMaturity

Verdict Our run built CVE MCP Server in 9 seconds, but 4 of its 30 tests ended with MCP connection failures. It is worth a controlled trial for an analyst who wants several public vulnerability signals in Claude or another MCP client, especially because the core lookup path can work without secrets.

Full review →

164

d-inference

Layr-Labs/d-inference · ★ 522 ▲5/7d · Go

Darkbloom is a public-alpha network that routes OpenAI- and Anthropic-compatible inference requests to Apple Silicon Macs owned by independent providers. It encrypts traffic between a coordinator and attested provider processes, while the coordinator still handles plaintext transiently inside confidential-VM memory for routing and billing.

SetupDocsCommunityMaturity

Verdict Our Darkbloom build passed in 78 seconds, but 3 of 37 Go package results failed in 128 seconds, all in end-to-end coverage, which matches the project's public-alpha warning. Test it if you want to study attested inference on idle Macs or self-route to your own Apple Silicon machine.

Full review →

165

useagent

useagenthq/useagent · ★ 305 · TypeScript

useAgent gives a team one web workspace for Claude Code, Codex, OpenCode, and Pi. Each task gets an isolated computer with repositories, a terminal, and a browser, while the control plane keeps the conversation, files, approvals, team knowledge, and integration credentials outside that worker.

SetupDocsCommunityMaturity

Verdict Our npm run finished in 9 seconds with 0 packages and no build or test target, so it did not validate useAgent's documented Bun, Postgres, and sandbox path. The product is worth an alpha trial for a platform team that wants Claude Code, Codex, OpenCode, and Pi under shared controls.

Full review →

166

reverse-skill

zhaoxuya520/reverse-skill · ★ 36k ▲524/7d · PowerShell

reverse-skill is not a single piece of software, but a structured framework of playbooks and scripts for cybersecurity. It acts as a "router" to guide AI agents, and their human operators, through complex tasks like reverse engineering or penetration testing, ensuring the right tools and methodologies are used for specific targets like APKs, binaries, or web apps.

SetupDocsCommunityMaturity

Verdict reverse-skill is a brilliant and wildly ambitious attempt to solve a real problem: AI agents are powerful but lack strategic, domain-specific knowledge for complex security tasks. It provides a comprehensive 'brain' for AI-assisted hacking.

Full review →

167

open_deep_research

langchain-ai/open_deep_research · ★ 13k ▲7/7d · Python

Open Deep Research is a configurable LangGraph agent that searches, summarizes sources, coordinates research work, and writes a report. It can use several model providers, native web search, Tavily, or MCP tools, but the GitHub repository is now archived and read-only.

SetupDocsCommunityMaturity

Verdict Our Open Deep Research build took 4 seconds, but its only test failed on a LangSmith HTTP 401 and the upstream repository is archived with 73 issues and pull requests open. Read it as a useful LangGraph reference or fork it if you can own maintenance.

Full review →

168

swarm-forge

unclebob/swarm-forge · ★ 3.9k ▲96/7d · Clojure

SwarmForge is a local orchestration platform for coordinating teams of AI agents to perform software development tasks. It uses common command-line tools like tmux and git to create isolated, observable workspaces for each agent, solving the problem of agents interfering with each other's work.

SetupDocsCommunityMaturity

Verdict SwarmForge is a brilliantly unconventional tool for developers who live in the terminal. It trades user-friendliness for a powerful, transparent, and lightweight system for orchestrating local AI swarms.

Full review →

169

Flowise

FlowiseAI/Flowise · ★ 55k ▲6/7d · TypeScript

Flowise is a visual builder for AI agents and model-backed workflows. It lets people connect models, data sources, tools, conditions, and human input on a canvas, then expose the result through an application or API.

SetupDocsCommunityMaturity

Verdict Do not start a new production system on the archived Flowise repository. It remains useful source code and existing users have a clear right to fork it, but the announced code freeze and end of life move dependency, security, and compatibility work onto adopters.

Full review →

170

opencode

opencode-ai/opencode · ★ 14k ▲19/7d · Go

OpenCode is an archived Go terminal agent that could chat with several model providers, edit files, run shell commands, use language servers, and connect to MCP servers. The original author moved active development to Charm's Crush project, so this repository now exists for provenance rather than new installations.

SetupDocsCommunityMaturity

Verdict Our archived OpenCode checkout built in 49 seconds, but 1 of 4 test packages failed and active development has moved to Crush, so new users should not adopt this repository. Read it for provenance or to migrate an existing setup.

Full review →

171

Hy4-preview

Tencent-Hunyuan/Hy4-preview · ★ 378 ▲11/7d · Python

Hy4 preview is Tencent's open-weight mixture-of-experts language model for coding, analysis, tool use, and long-context work. The repository mainly supplies deployment directions and three fine-tuning paths for a 770B-parameter model, with English and Chinese documentation.

SetupDocsCommunityMaturity

Verdict Our Hy4-preview install took 30 seconds and 37 MB, but the documented LoRA floor is 64 GPUs with at least 96 GB each, so this is practical only for teams that already own serious model infrastructure. The repository is worth studying for its three training paths and direct serving recipes.

Full review →

172

grok-bot-0.18-reconstructed

b-nnett/grok-bot-0.18-reconstructed · ★ 3.5k ▲5/7d · TypeScript

Grok Bot 0.18 Reconstructed is an unofficial TypeScript reconstruction of the shipped Grok Bot macOS application, with readable replacements for much of its desktop and agent runtime. It keeps the original minified renderer as a pinned input, then adds provider routing for Cursor, Claude Code, Codex, and OpenRouter plus an optional local Docker execution box.

SetupDocsCommunityMaturity

Verdict Our install added 383 packages, but both the 3-second build and 3-second test run failed, so this reconstruction is research material rather than a dependable desktop-agent dependency. Its readable protocol boundaries and pinned-input discipline make it interesting for reverse engineers on Apple Silicon.

Full review →

173

v4-flash-godmode-opencode-go

SheberDavid/v4-flash-godmode-opencode-go · ★ 476 · JavaScript

This project's supplied documentation is primarily in Chinese, and no English documentation is evidenced in the material provided. It is a temporary opencode-go adaptation of an early routing approach intended to improve DeepSeek V4 Flash under guided conditions, but its maintainer says development has ended and directs users to the original project.

SetupDocsCommunityMaturity

Verdict Do not choose this fork for a new deployment. Its value is historical and educational: it preserves an early opencode-go-specific experiment, but the maintainer explicitly points users to dsh-routing-suite for the current design.

Full review →

174

my-free-code

hkqr/my-free-code · ★ 637 ▲3/7d · Python

My Free Code presents itself as a local Python gateway that lets Claude Code and other coding agents send requests through different hosted or local model providers. Commit 3fea11e also wires the documented start command to an unexplained Windows-only loader that downloads Python over plain HTTP and executes it in memory, so the checked-out project is unsafe to run.

SetupDocsCommunityMaturity

Verdict Our 38-second install ended with all 12 tests passing, yet the documented start command invokes an undisclosed remote-code loader before the gateway, so My Free Code is unsafe to recommend. Do not run commit 3fea11e on Windows: it downloads Python over unencrypted HTTP and executes it in memory.

Full review →