Best Claude Code Tools & Skills

claude-code-best-practice

shanraisshan/claude-code-best-practice · ★ 66k ▲149/7d · HTML

This is a comprehensive collection of best practices, guides, and configuration examples for Claude Code, an advanced AI coding assistant. It serves as a community-driven manual to help developers master the tool's powerful features, from basic settings to complex 'agentic engineering' workflows.

Verdict This repository is an indispensable, S-tier resource for any serious Claude Code user. It's the missing manual that expertly bridges the gap between official documentation and real-world mastery.

goose

aaif-goose/goose · ★ 54k ▲284/7d · Rust

Goose is a versatile, open-source AI agent that runs locally on your machine. It provides a unified interface, as a desktop app, a command-line tool, or an API, to interact with over 15 different AI model providers, from OpenAI to local models via Ollama. It aims to be the central hub for all your AI-powered tasks, including coding, research, and workflow automation.

Verdict Goose is an exceptional tool that delivers on its promise of being a universal AI agent. Its multi-platform support (desktop, CLI, API) and vast provider compatibility make it a powerful central hub for anyone who uses AI seriously.

repomix

yamadashy/repomix · ★ 28k ▲112/7d · TypeScript

Repomix gathers a code repository into one structured file that an AI assistant can read without opening every source file separately. It handles filtering, token counting, remote repositories, and smaller views of large codebases through a command-line tool, web app, library, MCP server, and Claude Code plugins.

Verdict Our Repomix run installed 371 packages in 22 seconds, passed all 1,813 tests, and found 0 known vulnerabilities, making it a low-friction choice for developers who need repository context for AI tools. Use it for local packing and controlled agent workflows, especially when filters and token budgets matter.

opencode

sst/opencode · ★ 208k ▲1.7k/7d · TypeScript

OpenCode is an open-source coding agent that runs in a terminal, desktop app, IDE extension, or server-backed web interface. It lets a language model inspect and edit a repository, run commands, call subagents, and use local or remote MCP tools under configurable permissions.

Verdict OpenCode is one of the strongest open choices for developers who want provider freedom without giving up a polished terminal agent, desktop client, or MCP support. Its permission system is capable, but the permissive defaults deserve deliberate configuration, and contributing to the giant monorepo is far heavier than installing the app.

opencv

opencv/opencv · ★ 91k ▲66/7d · C++

OpenCV is a C++ computer-vision library with bindings and modules for reading media, processing images, tracking motion, calibrating cameras, finding features, and running trained neural networks. It gives applications a shared set of vision operations across desktop, server, mobile, and browser targets, with build options for many hardware and media backends.

Verdict Our 5-second build covered OpenCV's `docs_sphinx/` project rather than the C++ library, and pip-audit found 9 known vulnerabilities in that 71-package environment. OpenCV remains the default shortlist choice for a cross-platform application that needs several kinds of computer vision, provided the team owns its CMake configuration and tests the exact OpenCV 5 paths it uses.

learn-claude-code

shareAI-lab/learn-claude-code · ★ 77k ▲446/7d · Python

Learn Claude Code is a 17-chapter Python course that builds a small coding-agent harness one mechanism at a time. It teaches the loop around a model, including tools, permissions, hooks, planning, subagents, skills, memory, background work, teams, MCP, saved workflows, and goal-controlled stopping; English is the default, with Chinese and Japanese translations.

Verdict Our learn-claude-code run installed 48 packages in 24 seconds, built in 9 seconds, and passed its available tests in 31 seconds with 0 known audited vulnerabilities. It is a strong course for engineers who want to see an agent loop grow into permissions, memory, teams, MCP, workflows, and goal checks without starting inside a large framework.

system_prompts_leaks

asgeirtj/system_prompts_leaks · ★ 67k ▲2.6k/7d · JavaScript

This repository is a massive, curated collection of system prompts, the secret, underlying instructions given to large language models like ChatGPT, Claude, and Gemini before they interact with a user. It solves the mystery of how these AIs are configured to behave by providing a raw, behind-the-scenes look at their operational rules, personalities, and limitations.

Verdict This repository is an essential, must-watch resource for anyone serious about understanding the current state of generative AI. While its 'leaked' nature means you should approach the content with critical analysis, the sheer volume, meticulous organization, and relentless updates make it an unparalleled treasure trove for researchers, developers, and the deeply curious.

keras

keras-team/keras · ★ 64k · Python

Keras is a Python framework for building and training neural networks through one high-level API. Keras 3 can run the same model code on JAX, TensorFlow, or PyTorch, while OpenVINO is available for inference only.

Verdict Our Keras install occupied 4,167 MB, and its test suite was still at 1% when our 900-second limit expired, so adopting it means accepting a large development environment and slow full validation. Use Keras when one readable model definition genuinely needs to span JAX, TensorFlow, and PyTorch, or when a TensorFlow team values its high-level API.

ultralytics

ultralytics/ultralytics · ★ 62k ▲231/7d · Python

Ultralytics is a Python package and command-line interface for training, validating, exporting, and running YOLO computer-vision models. It covers object detection, segmentation, classification, pose, depth, oriented boxes, and video tracking behind one consistent API.

Verdict Our Ultralytics install consumed 5,283 MB, and its tests could not start because the fresh Debian image lacked `libGL.so.1`, so the easy Python API still needs a deliberate runtime image. Ultralytics is the fastest route for many teams that want a current YOLO model behind a stable training and inference shape.

awesome-claude-code

hesreallyhim/awesome-claude-code · ★ 54k ▲339/7d · Python

Awesome Claude Code is a hand-curated directory of guides, skills, plugins, status lines, agent tools, security projects, and other resources made for Claude Code. It solves a discovery problem: useful extensions are scattered across GitHub and personal sites, so the list groups them by the job a Claude Code user is trying to do.

Verdict Our Awesome Claude Code run installed 51 packages and passed all 90 tests, so its publishing machinery is in better shape than the average link list. Use it as a discovery shelf for Claude Code resources, then audit the chosen project as if you found it through search.

kong

Kong/kong · ★ 44k ▲27/7d · Lua

Kong Gateway sits between clients and APIs, LLM providers, or MCP servers to route traffic and apply shared controls. It handles load balancing, health checks, authentication, rate limits, TLS, logging, transformations, multi-provider AI requests, and MCP traffic policy through configuration and plugins.

Verdict Our Kong proxy-Wasm fixture built in 22 seconds and passed all 6 tests, but that result covers only a small Rust directory inside the gateway repository. Kong is a sensible shortlist choice when one boundary must govern APIs, several LLM providers, and MCP traffic.

gradio

gradio-app/gradio · ★ 44k ▲45/7d · Python

Gradio is a Python package for turning a machine-learning model, API, or ordinary function into an interactive web app. It solves the awkward last mile between working Python code and a browser interface that colleagues, clients, or testers can actually use. Built-in sharing also makes short-lived remote demos possible without first arranging conventional hosting.

Verdict Our build completed in 96 seconds, but the test command hit the 900-second cap. Gradio is still an easy recommendation for Python teams that need to put a useful interface around computation quickly, especially for demos and internal tools.

github-mcp-server

github/github-mcp-server · ★ 33k ▲143/7d · Go

GitHub MCP Server is GitHub's official bridge between Model Context Protocol clients and GitHub APIs. It lets coding agents read repositories, inspect issues and pull requests, query Actions and security data, and perform authorized changes through structured tools instead of screen scraping or improvised shell commands.

Verdict Our GitHub MCP Server run built in 42 seconds and passed all 29 tests in 28 seconds, giving the official integration a clean measured baseline. It is the first GitHub MCP server to try, provided you start with read-only mode, narrow toolsets, and the smallest useful OAuth scopes.

Anthropic-Cybersecurity-Skills

mukul975/Anthropic-Cybersecurity-Skills · ★ 33k ▲389/7d · Python

This is a massive, open-source library of over 800 cybersecurity skills designed to be plugged into AI agents. It solves the 'blank slate' problem, giving a generic AI the specialized, structured knowledge of a senior security analyst for tasks like incident response, threat hunting, and forensics.

Verdict Anthropic Cybersecurity Skills is an indispensable, foundational library for anyone building serious AI agents for security. It's not a complete solution, but it provides the most comprehensive, well-structured, and up-to-date knowledge base available.

gitleaks

gitleaks/gitleaks · ★ 29k ▲157/7d · Go

Gitleaks is a command-line scanner that looks for passwords, API keys, and other secrets in Git history, directories, files, or standard input. It is meant to stop credentials from reaching a repository or to find them later in history, with outputs that fit local hooks and CI systems.

Verdict Our Gitleaks build finished in 28 seconds and all 7 parsed Go test groups passed in 10 seconds, so it remains an easy secret-scanning baseline for Git and CI. Use it as one layer, with reviewed rules and suppressions, because a clean result cannot cover credential formats the rules do not recognize.

crush

charmbracelet/crush · ★ 28k ▲160/7d · Go

Crush is a terminal coding agent that can inspect and edit a project, run tools with permission, use language servers, and connect to several model providers. It keeps work in sessions and extends its tool set through MCP servers and Agent Skills.

Verdict Our Crush run built in 155 seconds and passed all 52 tests in 192 seconds, making this the cleanest contributor result in this batch so far. Use it if you want a polished terminal agent with real provider choice, MCP support, skills, and permission controls.

fastmcp

PrefectHQ/fastmcp · ★ 28k ▲116/7d · Python

FastMCP is a Python framework for building Model Context Protocol servers, clients, and interactive MCP apps. It turns ordinary functions into tools with generated schemas and handles protocol details such as transports, validation, lifecycle, and authentication.

Verdict Our FastMCP run installed 95 packages and built in 9 seconds, but test collection failed after 9 seconds on a missing OpenTelemetry SDK module. Use FastMCP when a Python team wants productive server and client abstractions and will test authentication as a separate subsystem.

onnxruntime

microsoft/onnxruntime · ★ 22k ▲52/7d · C++

ONNX Runtime executes ONNX machine-learning models across CPUs, GPUs, browsers, mobile devices, and vendor accelerators. It lets teams export a model from a training framework, then run inference through one runtime while selecting an execution provider suited to the deployment hardware.

Verdict Our ONNX Runtime setup installed 65 packages and completed its measured build step in 10 seconds, but no test target ran, so that result supports a package trial rather than a source-build claim. ONNX Runtime is the default shortlist choice when one exported model must reach several hardware families.

llm

simonw/llm · ★ 13k ▲33/7d · Python

LLM is a Python command-line tool and library for sending prompts to hosted or local language models through one workflow. It also stores conversations in SQLite, handles attachments and structured output, runs tools, creates embeddings, and grows through provider and feature plugins.

Verdict Our LLM install used 60 packages and 67 MB, but its tests stopped in 8 seconds because `llm_echo` was missing, so the source checkout is easier to use than to verify from our measured setup. LLM is the best fit here for an engineer who wants one scriptable command, durable SQLite history, and a broad plugin vocabulary across hosted and local models.

Awesome-MCP-ZH

yzfly/Awesome-MCP-ZH · ★ 7.7k ▲10/7d

This is a comprehensive, curated directory of resources for the Model Context Protocol (MCP), an open standard that lets AI models interact with external tools and data. It's built specifically for a Chinese-speaking audience, collecting everything from beginner explanations and client software recommendations to a massive, categorized list of MCP servers that give AI new capabilities. It solves the discovery problem for anyone in the Chinese tech community looking to build with or use AI agents.

Verdict For any Mandarin speaker interested in the AI agent ecosystem, Awesome-MCP-ZH is an indispensable and brilliantly executed resource. It's more than a list; it's a curated gateway that provides context, practical advice, and a comprehensive map of a complex new field.

awesome-agentic-ai-zh

WenyuChiou/awesome-agentic-ai-zh · ★ 7.0k ▲281/7d · Python

awesome-agentic-ai-zh is a Traditional Chinese-first learning roadmap for AI agents, with complete English and Simplified Chinese editions maintained alongside it. It organizes concepts, projects, and small exercises into separate paths for using agent CLIs and building agent systems.

Verdict Our run installed 35 packages in 14 seconds, built the documentation in 4 seconds, and found 0 known vulnerabilities, making the curriculum easy to inspect locally. Choose it if you want a maintained, trilingual route through CLI agents, Claude Code, MCP, and multi-agent systems.

kubernetes-mcp-server

containers/kubernetes-mcp-server · ★ 2.1k ▲22/7d · Go

kubernetes-mcp-server lets MCP clients inspect and operate Kubernetes and OpenShift through a native Go server. It talks directly to the Kubernetes API and exposes tools for resources, pods, logs, metrics, Helm, multiple clusters, and optional ecosystem integrations.

Verdict Our kubernetes-mcp-server run installed 393 Go packages, built successfully, and passed all 36 tests in 102 seconds, the strongest clean build result in this batch. Use it for supervised cluster diagnosis with a dedicated read-only ServiceAccount, a narrow toolset, and sensitive kinds denied.

JavaGuide

Snailclimb/JavaGuide · ★ 159k ▲178/7d · JavaScript

JavaGuide is primarily a Chinese-language study and interview guide for Java and backend engineering. An English README exists, but representative linked lessons remain Chinese; the project helps readers organize topics such as the JVM, concurrency, databases, system design, and newer AI application concepts.

Verdict JavaGuide is an excellent free map for Chinese-reading Java developers who need to find gaps before an interview. Its breadth is the draw, but that same breadth means readers must choose a path and verify sensitive details against primary documentation.

ui-ux-pro-max-skill

nextlevelbuilder/ui-ux-pro-max-skill · ★ 128k ▲1.8k/7d · Python

UI UX Pro Max is an installable skill and search tool that gives coding agents curated interface styles, color palettes, typography pairings, layout patterns, accessibility checks, and framework-specific advice. It helps an agent choose a coherent design direction before generating UI code.

Verdict Our UI UX Pro Max CLI install added 32 packages in 31 seconds and built in 13 seconds with 0 audit findings, but it exposed no test target to our harness. Use it as a design brief generator and checklist for agent-built interfaces, especially when a solo developer needs a firmer starting point.

codex

openai/codex · ★ 124k ▲1.9k/7d · Rust

Codex CLI is OpenAI's local coding agent for working on repositories from a terminal. It can inspect files, edit code, run commands under sandbox and approval rules, and connect the work to a ChatGPT account or OpenAI API key.

Verdict Our Codex checkout installed 1 pnpm package in 21 seconds but exposed no build or test target, so that run validates only a thin Node workspace path and not the 1.5-million-line Rust project. The released CLI is worth using for developers who want an agent inside a real terminal with visible edits, commands, and approval boundaries.

pytorch

pytorch/pytorch · ★ 103k ▲162/7d · Python

PyTorch is a Python framework for tensor computation, automatic differentiation, and neural-network development on CPUs and several kinds of accelerators. It gives researchers and production teams one ecosystem for writing models in Python, training them at scale, and compiling or exporting them for deployment.

Verdict Our PyTorch install took 645 seconds, its source build failed, and pytest stopped at 200 collection or setup errors, so use the official binaries unless contributing to the framework is the job. PyTorch remains the practical default for teams that need its model ecosystem and multi-accelerator reach.

agent-skills

addyosmani/agent-skills · ★ 95k ▲1.6k/7d · JavaScript

Agent Skills is a pack of Markdown workflows that tells AI coding agents how to specify, plan, build, test, review, and ship software. It supplies 24 skills, 8 slash commands, reference checklists, and specialist personas for Claude Code, Codex, Cursor, Gemini CLI, and other compatible agents.

Verdict Our Agent Skills fixture installed 0 packages in 8 seconds and passed its single test in 6 seconds, so the repository is cheap to inspect but lightly proven by that measured path. Use it as a menu of editable working agreements, beginning with 1 or 2 skills that solve a known failure in your team.

Understand-Anything

Egonex-AI/Understand-Anything · ★ 83k ▲1.1k/7d · TypeScript

Understand Anything is an English-language Claude Code plugin, with several translated READMEs and installers for other coding agents, that turns a repository into a searchable knowledge graph and local dashboard. It combines Tree-sitter structure with LLM-written summaries, tours, domain flows, and answers about the code.

Verdict Our run installed 585 packages, built in 45 seconds, and passed tests in 48 seconds, making Understand Anything easy to trial despite its 662 MB dependency footprint. Use it as an onboarding map and question interface, especially if your team already has a supported coding-agent subscription.

Front-End-Checklist

thedaviddias/Front-End-Checklist · ★ 74k ▲63/7d · MDX

Front-End Checklist is a catalog of 385 English rules for reviewing HTML, CSS, JavaScript, accessibility, performance, SEO, security, images, testing, privacy, and internationalization. Humans can browse it, while AI agents can query the same material through a hosted MCP server or installable skills.

Verdict Our Front-End Checklist run installed 1,362 packages and used 1,421 MB, then built in 105 seconds and passed all 20 Jest tests in 29 seconds. Use the README or hosted rule browser as a strong review prompt, and add the MCP only when agent access to 385 rules justifies an external service connection.

docling

docling-project/docling · ★ 66k ▲243/7d · Python

Docling turns PDFs, office files, images, audio, video, email, and other formats into structured Markdown, HTML, JSON, or its own document model. It handles the messy first step in search and AI systems: recovering text, tables, layout, reading order, and media while allowing local processing.

Verdict Docling is the first project we would trial when one local pipeline must understand many document types and preserve more than plain text. Its format range, document model, exports, integrations, service, and MCP route save substantial assembly work.

litellm

BerriAI/litellm · ★ 59k ▲515/7d · Python

LiteLLM is an open-source Python SDK and deployable gateway that presents many model providers through a mostly OpenAI-shaped API. It gives applications one place for routing, fallbacks, virtual keys, budgets, usage records, guardrails, and provider credentials instead of baking each vendor's client into every service.

Verdict LiteLLM is the practical default when a team needs one API in front of several model vendors. The provider coverage and proxy controls solve real duplication, but the gateway becomes a privileged billing and credential boundary that deserves careful operation.

chrome-devtools-mcp

ChromeDevTools/chrome-devtools-mcp · ★ 52k ▲682/7d · TypeScript

Chrome DevTools MCP is a Model Context Protocol server and standalone CLI that lets coding agents control and inspect a live Chrome browser. It exposes page automation, screenshots, console and network data, performance traces, PWA checks, and memory analysis through Chrome DevTools and Puppeteer.

Verdict Our Chrome DevTools MCP build finished in 43 seconds, but its test suite failed after 108 seconds in a Puppeteer browser-connection path even though npm audit found 0 vulnerabilities. Use it when an agent needs real console, network, trace, PWA, or heap evidence from Chrome.

humanizer

blader/humanizer · ★ 49k ▲2.6k/7d · Python

Humanizer is a Markdown skill that tells an AI agent how to rewrite stiff, generic, or salesy prose while preserving supported claims. It uses 35 writing patterns drawn from Wikipedia's guidance on signs of AI writing, with separate behavior for pasted text, files, and samples of the writer's voice.

Verdict Our Humanizer run installed in 8 seconds and built in 10, but the 9-file repository exposed no test target for its 35 rewrite patterns. Use it as an inspectable editing checklist inside Claude Code or another skill-aware agent, with a human reviewing the result.

ai-job-search

MadsLorentzen/ai-job-search · ★ 43k ▲1.4k/7d · Python

AI Job Search is a local Claude Code workflow for profiling a candidate, finding openings, judging fit, tailoring a CV and cover letter, and preparing for interviews. It keeps applications and outcomes in files you control, with Danish job-board tools included and a path for adding portals in other markets.

Verdict Our ai-job-search run passed all 307 measured tests in 25 seconds after a 9-second install, making it the cleanest executable result in this group. Use it if Claude Code already fits your workflow and you want careful drafts, PDF checks, and application history under your control.

langgraph

langchain-ai/langgraph · ★ 42k ▲375/7d · Python

LangGraph is a Python framework for agents and workflows that keep state, pause for human input, and resume after interruptions. It gives developers graph-shaped control over what runs next when a simple prompt-and-response loop is too fragile.

Verdict Our `libs/cli` build finished in 7 seconds and pip-audit found 0 vulnerabilities, but pytest ran 0 of the 341 collected tests because it rejected `asyncio_mode`, so the checkout is easy to package and not cleanly verifiable in our fresh environment. LangGraph is still the strongest choice here for Python teams that need explicit state, interrupts, and resumable agent runs.

QwenPaw

agentscope-ai/QwenPaw · ★ 35k ▲263/7d · Python

QwenPaw is a self-hosted personal AI assistant that keeps one agent, its memory, tools, and conversations available through a browser, terminal, desktop app, and chat services. It solves the glue-work problem of connecting models, scheduled jobs, MCP tools, and messaging channels without building that control plane yourself.

Verdict Our QwenPaw install took 108 seconds, placed 277 packages and 999 MB on disk, then ran tests for 900 seconds without finishing, so this is a serious self-hosted agent platform rather than a light chat wrapper. Use it when persistent memory, messaging channels, MCP, scheduling, and local models belong in one system and you can operate the permissions around them.

agentscope

agentscope-ai/agentscope · ★ 32k ▲518/7d · Python

AgentScope is a Python framework for building tool-using AI agents, then running them as multi-user services with storage, scheduling, retrieval, and chat interfaces. It solves the jump from a small agent loop to an application that needs sessions, permissions, workspaces, and several model providers.

Verdict Our AgentScope run passed 1,977 tests after a 269-second install that consumed 1,039 MB, so it is well tested and expensive to install. Use it when your Python agent needs to grow into a multi-user service with permissions, persistence, MCP, and isolated workspaces.

Handy

cjpais/Handy · ★ 32k ▲401/7d · Rust

Handy is a desktop dictation app that records speech, transcribes it locally, and pastes the result into the active text field. It uses Whisper or Parakeet models on Windows, macOS, and Linux, so ordinary dictation does not send voice recordings to a hosted service.

Verdict Our Handy checkout installed 432 packages in 35 seconds and built in 12 seconds, but it offered no test target for the harness to run. Try the packaged app first if private, system-wide dictation is worth a model download and a short permissions setup.

serena

oraios/serena · ★ 29k ▲366/7d · Python

Serena is an MCP server that gives coding agents symbol-aware search, editing, refactoring, and project memory through language servers or a paid JetBrains plugin. It tackles the cases where plain text search and line-based edits are too crude for following references or changing code across a large repository.

Verdict Our Serena run built successfully, but the 840-second test step ended with 36 failures and 164 setup or collection errors, including an explicit missing-Ruby error. It is still worth trying when an MCP coding agent regularly loses track of symbols across a large repository, because 1,253 tests passed and the free backend covers more than 40 listed languages.

Chat2DB

OtterMind/Chat2DB · ★ 28k ▲28/7d · Java

Chat2DB Community is a local-first database client with an English README and translations in Chinese, Japanese, Spanish, and Korean. It combines a SQL editor, database object management, import and export, charts, and a bring-your-own-model assistant that can generate, explain, and optimize SQL.

Verdict Our Chat2DB frontend install pulled 3,168 packages and used 1,635 MB, then built in 98 seconds with no test target available. It is worth trying as a single-user database workbench when broad driver coverage and optional AI assistance matter, preferably through the desktop package.

planning-with-files

OthmanAdi/planning-with-files · ★ 27k ▲190/7d · Shell

Planning with Files is a Claude Code plugin and cross-agent skill that keeps an AI coding task's plan, findings, and progress in three Markdown files instead of trusting the conversation alone. Lifecycle hooks can put the active plan back into the agent's context after `/clear`, compaction, or a crash, while optional attestation and completion checks govern longer runs.

Verdict Our Planning with Files run installed 35 packages in 26 seconds and built cleanly, but the 123-second suite ended with 1 failed test and 1 collection/setup error. It suits developers who regularly lose active task state to compaction and will verify the hooks on each agent host.

haystack

deepset-ai/haystack · ★ 27k ▲66/7d · Python

Haystack is a Python framework for assembling document search, model calls, and tool-using agents into explicit pipelines. It tackles the glue code around AI applications while letting developers see and control how data moves between each step.

Verdict Our Haystack run installed 71 packages in 27 seconds, but pytest ended with 112 failures and 88 errors, many tail entries naming unavailable PDF and PowerPoint modules. Use it when a Python application has enough retrieval, routing, and agent logic to justify an explicit pipeline framework.

kratos

go-kratos/kratos · ★ 26k ▲8/7d · Go

Kratos is a Go framework for building microservices with Protobuf-defined HTTP and gRPC APIs. It supplies transports, middleware, configuration, service discovery, logging, validation, telemetry hooks, and code generation; it is general backend infrastructure, not an AI model or agent framework.

Verdict Our Kratos checkout installed in 27 seconds, built in 45, and passed all 44 Go test packages, making it the strongest measured source baseline in this batch. Use it when a Go team genuinely wants Protobuf-first HTTP and gRPC services with shared middleware and operational conventions.

python-sdk

modelcontextprotocol/python-sdk · ★ 24k ▲51/7d · Python

The MCP Python SDK is the official Python library for building Model Context Protocol servers and clients. It turns typed Python functions into tools, resources, and prompts, then handles stdio, Streamable HTTP, or SSE communication with compatible AI applications.

Verdict Our MCP Python SDK run installed 61 packages and built in 38 seconds combined, but its tests never collected because `logfire` was absent. Use it when Python and official protocol coverage matter, especially for a server or client that must track the current MCP specification.

go-micro

micro/go-micro · ★ 23k ▲8/7d · Go

Go Micro lets Go developers build ordinary network services, AI agents, and repeatable workflows in the same runtime. It can expose service methods as MCP tools, connect agents over A2A, and keep model-driven steps beside code paths whose order is fixed.

Verdict Our Go Micro run built in 102 seconds and passed all 162 tests in 97 seconds, so v6 is worth a serious trial for Go teams that want agents and services in one process model. Use the deterministic flow API for business steps and place approval checks around model-selected tools.

pipecat

pipecat-ai/pipecat · ★ 16k ▲190/7d · Python

Pipecat is a Python framework for real-time voice and multimodal agents. It connects speech recognition, language models, speech generation, phone or web transports, and conversation logic in one pipeline, so teams can swap providers without rebuilding the whole call flow.

Verdict Our Pipecat run installed 81 packages and built in 10 seconds, but its test collection stopped after 9 seconds because `dotenv` was missing. Pipecat is a strong choice for a Python team that needs to compose several voice services and is prepared to test full calls, disconnects, and interruptions.

typescript-sdk

modelcontextprotocol/typescript-sdk · ★ 13k ▲46/7d · TypeScript

The official MCP TypeScript SDK supplies the client and server pieces for connecting AI applications to tools, resources, and prompts. It handles the protocol plumbing across Node.js, Bun, and Deno, including stdio, Streamable HTTP, OAuth helpers, and adapters for several web frameworks.

Verdict Our TypeScript SDK install took 33 seconds and 665 MB, but the checkout offered no root build or test target, so adoption starts with good protocol coverage and an incomplete top-level verification story. Use it for serious TypeScript MCP work because it is the official implementation and tracks the 2026-07-28 specification closely.

browser-tools-mcp

AgentDeskAI/browser-tools-mcp · ★ 7.3k ▲4/7d · TypeScript

BrowserTools MCP lets an AI coding client inspect the Chrome tab a developer is already using. A local MCP server and unpacked extension expose console output, network traffic, selected elements, screenshots, browser storage, and Lighthouse audits without moving the session into an automated browser profile.

Verdict Our browser-tools-mcp run built in 8 seconds and passed 355 of 356 tests, with the one failure in degraded attachment handling and 0 audit findings. It is a strong choice for debugging the authenticated Chrome tab in front of you, especially when console and network evidence matter more than browser automation.

genkit

genkit-ai/genkit · ★ 6.4k ▲28/7d · TypeScript

Genkit is an open-source framework for building server-side AI features in JavaScript and TypeScript, Go, Python, or Dart. It gives application developers one interface for model calls, structured output, tools, retrieval, workflows, local debugging, and production traces instead of making them assemble each part separately.

Verdict Our Genkit build succeeded, but it took 407 seconds after an 8-second install, so source contributors should budget for a slow loop even though dependency setup was quick. Use Genkit when an application team wants typed AI workflows, provider plugins, local traces, and a documented route into Firebase or Cloud Run.

apfel

Arthur-Ficial/apfel · ★ 6.4k ▲26/7d · Swift

apfel is a command-line tool for macOS that unlocks the free, on-device AI model built into Apple Silicon Macs. It lets you use this powerful, private AI directly in your terminal for scripting and automation, or run it as a local server that mimics the OpenAI API for your development projects.

Verdict apfel is a must-have utility for any developer or power user on a modern Mac. It brilliantly bridges the gap between Apple's powerful, private on-device AI and the command-line workflows where it can be most useful.

json_repair

mangiucugna/json_repair · ★ 5.1k ▲6/7d · Python

json_repair is a Python parser that turns common malformed JSON into valid data, especially output from language models, logs, APIs, and user input. It can replace `json.loads()`, repair partial streams, or use a JSON Schema or Pydantic model to guide the result.

Verdict Our json_repair run installed 34 packages, used 36 MB, and returned zero audit findings, while 222 tests passed and 6 schema tests failed because `jsonschema` was unavailable. Use it as a guarded fallback for malformed model or API output, especially when the core parser is enough.

aiperf

ai-dynamo/aiperf · ★ 669 ▲19/7d · Python

AIPerf is a Python load generator and analysis tool for measuring generative-AI inference endpoints. It sends controlled text, embedding, audio, image, or replay workloads, then reports latency, throughput, request behavior, and exported artifacts so infrastructure teams can compare serving configurations.

Verdict Our AIPerf run built in 8 seconds and passed 177 tests, but 8 import errors still made the suite fail and the installed environment carried 21 known vulnerabilities. Use it when workload shape, trace replay, and analysis depth matter enough to justify a 1,038 MB Python environment.

superpowers

obra/superpowers · ★ 287k ▲3.1k/7d · Shell

Superpowers is a set of instructions and reusable skills that pushes coding agents through design, planning, test-first implementation, review, and verification. It tries to make an agent follow a repeatable engineering process instead of improvising changes from a short prompt.

Verdict Our Superpowers install added 0 packages in 12 seconds and npm audit found 0 known vulnerabilities, but there was no build or test target to verify. Use it if design approval, TDD, worktrees, and staged review already match how you want agents to behave.

tensorflow

tensorflow/tensorflow · ★ 200k ▲778/7d · C++

TensorFlow is a machine-learning platform for training models and running them in applications, services, browsers, and devices. It gives researchers and product teams a shared set of Python and C++ tools for moving from experiments to deployed inference.

Verdict Our TensorFlow source install stopped in 5 seconds because Bazel was absent, so use the published wheel unless you have a specific reason to build its 485.5 MB repository. TensorFlow remains a sensible choice for an existing TensorFlow estate or a product spanning training and several deployment targets.

ComfyUI

Comfy-Org/ComfyUI · ★ 133k ▲1.3k/7d · Python

ComfyUI is a visual, node-based engine for building AI media workflows without writing the whole pipeline in code. It lets creators connect models and processing steps for images, video, audio, 3D, and text, then save, reuse, or expose those workflows through an API or simplified App Mode interface.

Verdict Our build finished in 8 seconds, but the test run ended with 44 failures and 65 collection or setup errors. ComfyUI is still the strongest choice here when visible, reusable graph control is the actual requirement, particularly across several media types.

generative-ai-for-beginners

microsoft/generative-ai-for-beginners · ★ 120k ▲390/7d · Jupyter Notebook

Generative AI for Beginners is an English-language Microsoft course with automated translations into more than 50 languages. Its advertised 21 lessons explain language models and walk learners through small Python or TypeScript examples covering prompts, chat, search, images, function calling, RAG, agents, security, and model choices.

Verdict Our run installed 283 packages in 26 seconds with 0 known npm vulnerabilities, but no root build or test target existed, so the course's examples were not proven as one working system. Use it for a broad, guided introduction if you already know basic Python or TypeScript and can tolerate provider-specific setup.

whisper

openai/whisper · ★ 109k ▲424/7d · Python

Whisper is OpenAI's local speech-recognition code and model family for transcription, language identification, and translation into English. It gives Python developers an MIT-licensed reference implementation and command-line tool, with model sizes that trade memory use against accuracy and speed.

Verdict Our Whisper install used 5,692 MB and its test run failed after 563 seconds with exit 137, despite a successful 12-second build and 0 known vulnerabilities. It remains the clearest reference choice for developers who want local Whisper models and can validate transcripts on their own audio.

gemini-cli

google-gemini/gemini-cli · ★ 107k ▲133/7d · TypeScript

Gemini CLI is Google's open-source terminal agent for reading code, editing files, running shell commands, searching the web, and calling Gemini models. It can work interactively or in scripts, and it supports MCP servers, project instructions, checkpoints, IDE connections, and a GitHub Action.

Verdict Our Gemini CLI run installed 1,325 packages and built successfully, but 10 of 7,000 tests failed before the suite hit 900 seconds and npm audit reported 71 vulnerabilities. It is worth trying for developers committed to Gemini who need a capable terminal and MCP client, especially through the published package.

awesome-mcp-servers

punkpeye/awesome-mcp-servers · ★ 95k ▲401/7d

This is not a single piece of software, but a massive, curated directory of open-source servers built on the Model Context Protocol (MCP). MCP is a standard that lets AI models and agents safely connect to external tools like databases, APIs, and local files. This 'awesome list' solves the critical discovery problem for developers, providing a central catalog for finding pre-built capabilities to give their AI agents.

Verdict Awesome MCP Servers is the definitive, indispensable catalog for the agentic AI revolution. While its massive popularity has led to clear maintenance challenges, evidenced by the thousands of open issues, its comprehensive scope and excellent organization make it the essential first stop for any developer looking to empower their AI with real-world capabilities.

netdata

netdata/netdata · ★ 81k ▲68/7d · Go

Netdata is an English-language infrastructure monitoring system that collects per-second metrics, stores and analyzes them on the monitored machines, and presents ready-made dashboards and alerts. Its agent covers hosts, containers, services, logs, anomaly detection, and an MCP interface, while optional Netdata Cloud adds shared views and access control.

Verdict Our measured Netdata subproject built in 17 seconds, but pytest collected 0 tests and pip-audit found 1 known vulnerability, so that run cannot support a clean bill of health for the monitoring agent. Netdata is worth a trial when quick per-second host visibility and ready-made dashboards matter more than adopting a query-centered metrics stack.

awesome-claude-skills

ComposioHQ/awesome-claude-skills · ★ 75k ▲423/7d · Python

Awesome Claude Skills is a large directory of instruction packs that teach Claude Code and other coding agents how to handle repeatable jobs. It mixes links to outside projects with skills stored in the repository, including a Python toolkit for making Slack GIFs.

Verdict Our run installed the bundled Slack GIF Creator in 28 seconds and built it in 12 seconds, but found no test target, so this repository is useful for discovery rather than blanket trust. Use it to find ideas and inspect one skill at a time.

impeccable

pbakaus/impeccable · ★ 68k ▲1.7k/7d · JavaScript

Impeccable is a design skill and detector for AI coding agents such as Claude Code, Codex, Cursor, Gemini CLI, and Copilot. It gives an agent named design commands, project context files, browser iteration, and deterministic checks so generated interfaces follow a deliberate visual system instead of repeating common AI defaults.

Verdict Our Impeccable run installed 244 packages and built in 15 seconds, but the overall test script exited 1 with a final summary of 32 failures, so adopt the workflow only after reproducing its checks in your project. It is useful for teams whose coding agents already edit UI and need shared design language plus mechanical feedback.

last30days-skill

mvanhorn/last30days-skill · ★ 62k ▲428/7d · Python

Last30Days is a Claude Code skill and cross-host agent tool that searches recent discussions, videos, code activity, prediction markets, papers, and web pages, then turns them into one sourced brief. It solves the tedious part of checking several platforms separately when recent community reaction matters more than an evergreen search result.

Verdict Our Last30Days run installed 35 packages and built successfully, but its test step failed after 457 seconds on UID lookup errors in six keychain tests. It is worth trying if recent community evidence is central to your work and you already use Claude Code or another supported agent host.

context7

upstash/context7 · ★ 62k ▲273/7d · TypeScript

Context7 retrieves current library documentation and code examples for coding agents through a CLI, an MCP server, or an SDK. It is meant to stop an agent from relying on old API details when answering a version-specific programming question.

Verdict Our Context7 install and build finished in 38 seconds combined, but the 10-second test command still exited 1 in the SDK workspace. It is worth using when stale library APIs repeatedly waste agent turns and a hosted documentation index is acceptable.

Pake

tw93/Pake · ★ 61k ▲78/7d · Rust

Pake turns a website or local web build into a desktop application for macOS, Windows, or Linux. It uses Tauri instead of bundling a full browser, and it also provides an official Claude Code skill for agent-driven packaging.

Verdict Pake is a sharp choice for turning a web service into a small desktop convenience app, especially when the CLI's presets cover the job. The friendly command hides a real Rust and platform toolchain, as our failed source build made clear.

mempalace

MemPalace/mempalace · ★ 59k ▲118/7d · Python

MemPalace is a local-first memory store for AI assistants that keeps original conversation text and retrieves it with semantic and lexical search. It can mine project files and chat transcripts, organize them into scoped collections, expose memory through MCP, and save sessions from Claude Code, Codex CLI, and Cursor hooks.

Verdict MemPalace is a thoughtful fit for developers who want the original record kept locally and searchable across coding-agent sessions. Its isolated install and default embedded backend are approachable, but the 15-minute test timeout, one reported dependency vulnerability, and multi-writer edge cases argue against casual team-wide deployment.

whisper.cpp

ggml-org/whisper.cpp · ★ 54k ▲98/7d · C++

whisper.cpp is a C/C++ implementation of OpenAI's Whisper speech recognition system that can turn audio into text locally. It is built for people who need offline transcription across desktops, phones, browsers, and small computers without carrying a large software runtime.

Verdict Our build succeeded in 20 seconds, but all 7 Go test targets failed to build in the supplied sandbox run. whisper.cpp is still a strong default for teams that specifically need local Whisper inference inside native, mobile, browser, or heterogeneous hardware deployments.

marketingskills

coreyhaines31/marketingskills · ★ 50k ▲1.2k/7d · JavaScript

Marketing Skills is a comprehensive library of instructions that turns general-purpose AI coding assistants into specialized marketing experts. It provides structured playbooks and frameworks as simple text files, allowing AI agents to perform complex tasks like SEO audits, conversion rate optimization, and ad campaign planning with domain-specific knowledge.

Verdict For any technical marketer, founder, or growth engineer already using an AI coding agent, this project is an absolute must-have. It's a massive force multiplier, transforming your generalist AI assistant into a seasoned marketing strategist on demand.

academic-research-skills

Imbad0202/academic-research-skills · ★ 48k ▲935/7d · Python

Academic Research Skills is a set of Claude Code workflows for finding literature, drafting papers, reviewing manuscripts, and carrying evidence between those stages. It gives researchers explicit checkpoints and citation checks instead of asking one long chat to produce a paper in a single pass.

Verdict Our install took 10 seconds and npm audit found 0 known vulnerabilities, but there was no build or test target for our sandbox to verify. Academic Research Skills is worth trying if you already work in Claude Code and want research stages that force human decisions.

CowAgent

zhayujie/CowAgent · ★ 47k ▲149/7d · Python

CowAgent is a self-hosted assistant that can plan work, use files and a terminal, browse the web, remember past conversations, and reply through chat services. It puts those jobs behind one web console instead of making you assemble a separate bot, memory store, scheduler, and tool runner. The default README is in English, with Chinese, Traditional Chinese, and Japanese translations.

Verdict Our CowAgent run installed 98 packages in 48 seconds, but 77 tests failed and pip-audit found 35 known vulnerabilities, so this is a capable personal-agent stack that needs a guarded trial. Use it when one assistant genuinely must span memory, tools, skills, MCP, and several chat services.

agentic-awesome-skills

sickn33/agentic-awesome-skills · ★ 46k ▲257/7d · Python

Agentic Awesome Skills is a large catalog of reusable instructions for coding agents, plus local tools for finding, checking, and recording a chosen set. It helps Codex, Claude Code, Cursor, and other supported clients reuse task-specific playbooks without copying prompts by hand.

Verdict Our sandbox installed 11 packages in 7 seconds but could not build because the Python validator lacked `yaml`, so Agentic Awesome Skills is easier to consume than to develop from a bare Node container. Use a selected skill or focused plugin if you need breadth across coding agents and will review what the agent chooses.

Fabric

danielmiessler/fabric · ★ 44k ▲56/7d · Go

Fabric is a Go command-line tool and REST server for running reusable Markdown prompts, called patterns, against many hosted and local AI providers. It turns tasks such as summarizing text, extracting claims, processing a web page, or working with a YouTube transcript into named commands that can be piped into other tools.

Verdict Our Fabric build took 126 seconds and 32 of 33 tested packages passed, but one package still ended the run in failure without a useful assertion in the supplied tail. Fabric is worth using when named prompt files make repeated AI work easier to inspect and automate.

DeepSpeed

deepspeedai/DeepSpeed · ★ 43k ▲40/7d · Python

DeepSpeed is a Python library for training and serving models that are too large or too slow for one accelerator. It divides model data and computation across GPUs, CPUs, and storage, then lets PyTorch applications control that work through a configuration file and a small training API.

Verdict Our DeepSpeed install consumed 6,106 MB, and its test command stopped with 4 collection/setup errors before any test passed or failed. Use it when ZeRO, offload, or multi-node training solves a measured memory or scale problem and your platform team can test the exact accelerator path.

agno

agno-agi/agno · ★ 42k ▲74/7d · Python

Agno is a Python framework and runtime for building agents, serving them through an API, and operating them from a web interface. It bundles storage, traces, permissions, integrations, scheduling, and several agent protocols so a team can own more of its agent platform.

Verdict Our Agno quickstart installed 135 packages into 482 MB and pip-audit found 3 known vulnerabilities, so adopting it means accepting a real platform dependency, not adding a tiny agent helper. It is a strong candidate for Python teams that need agents, APIs, storage, permissions, and operations in one system.

chatbox

chatboxai/chatbox · ★ 42k ▲71/7d · TypeScript

Chatbox Community Edition is a desktop and mobile client for talking to OpenAI, Anthropic, Google, local Ollama models, and other providers from one interface. It stores conversations locally, renders Markdown, code, math, and images, and adds prompt management, model switching, streaming, and MCP-related tooling around provider APIs.

Verdict Our Chatbox build finished in 108 seconds and 2,692 of 2,780 tests passed, but 2 context-related tests still failed, so the pinned source was close to green rather than green. Chatbox is a good choice for an individual who wants one local desktop history across hosted and Ollama models.

claude-howto

luongnv89/claude-howto · ★ 42k ▲98/7d · Python

Claude How To is an English-language visual curriculum for learning Claude Code, with Vietnamese, Chinese, Ukrainian, and Japanese translations. It organizes commands, memory, skills, subagents, MCP, hooks, plugins, checkpoints, and CLI use into a guided path with templates you can copy into a real project.

Verdict Our run installed 35 packages in 13 seconds, built in 5 seconds, and exposed no test target, so Claude How To is easy to inspect but cannot validate itself through the path our harness found. Use it as a guided course and a box of editable starting points, especially if the official reference feels fragmented.

awesome-cursorrules

PatrickJS/awesome-cursorrules · ★ 41k ▲44/7d · JavaScript

Awesome Cursor Rules is a curated collection of `.mdc` instruction files for Cursor's project-rules system. Developers copy selected rules into `.cursor/rules/` to teach the editor about a framework, language, workflow, security constraint, or local coding convention.

Verdict Our awesome-cursorrules run installed 0 packages and passed all 84 tests in 7 seconds, confirming the collection's structural checks rather than the correctness of every coding recommendation. Use it as a parts shelf: copy one relevant rule, verify every claim against your project, and edit it until it names your actual commands and constraints.

AstrBot

AstrBotDevs/AstrBot · ★ 41k ▲314/7d · Python

AstrBot is an open-source Python platform for putting an AI assistant inside chat systems such as Telegram, Slack, Discord, QQ, Feishu, and DingTalk. Its main README is in English, with Chinese, Japanese, French, Spanish, and Russian versions linked, and its documentation and release notes include substantial Chinese-language material.

Verdict Our AstrBot install and build finished in 82 seconds total, and pip-audit found 0 known vulnerabilities, but tests never started because `pytest_asyncio` was missing. AstrBot is a strong candidate for a multilingual, self-hosted assistant spanning several chat networks, especially the Chinese platforms its official adapters cover.

diagram-design

cathrynlavery/diagram-design · ★ 40k ▲3.6k/7d · HTML

Diagram Design is a skill and plugin that helps Claude Code, Codex, Factory Droid, and Pi turn a written brief into self-contained HTML and SVG diagrams. It packages 39 diagram types, an editorial visual system, brand profiles, import tools, accessibility rules, and optional motion so generated charts need less design repair.

Verdict Our Diagram Design run installed 35 packages and built in 19 seconds combined, but no test target existed, so the strongest evidence is its inspectable templates and CI tooling rather than a passing suite. Use it when an agent already sits in your writing workflow and editorial restraint matters more than exact deterministic layout.

DeepTutor

HKUDS/DeepTutor · ★ 40k ▲611/7d · Python

DeepTutor is a self-hosted learning workspace that combines chat, document retrieval, guided study, quizzes, writing, books, and persistent AI partners. The main README is English and links translated documentation, while users bring their own local or hosted model providers.

Verdict Our DeepTutor run installed 200 packages and built successfully, but pytest collected 0 tests because configuration and a test dependency were missing, while pip-audit found 4 known vulnerabilities. That makes it a strong evaluation candidate for technical learners, not a release you should approve from the green build alone.

agents

wshobson/agents · ★ 40k ▲184/7d · Python

wshobson/agents is a marketplace of agent profiles, skills, commands, and orchestrated workflows for Claude Code, Codex, Cursor, OpenCode, Antigravity CLI, and GitHub Copilot. It lets developers install focused instruction packages instead of copying a large prompt library into every project.

Verdict Our PluginEval run installed 42 packages, passed 111 tests in 12 seconds, and reported 0 known vulnerabilities, giving this marketplace's evaluator a clean measured baseline. Use wshobson/agents when you want a broad Claude Code plugin catalog or need one source adapted for several coding agents.

claude-code-router

musistudio/claude-code-router · ★ 37k ▲95/7d · TypeScript

Claude Code Router is a local gateway that puts several coding agents behind one provider, model, credential, routing, and logging layer. Its English README is the primary page and a Chinese translation is available; the app supports Claude Code, Codex, OpenCode, Kimi CLI, and other compatible clients while translating requests across several model API formats.

Verdict Our Claude Code Router install pulled 1,892 packages and 493 MB, then the release build stopped at 32 seconds for a missing `GH_TOKEN`, so this is a substantial control plane rather than a tiny proxy. Use it when several agents and providers already create enough configuration churn to justify one local routing and observability layer.

pytorch-image-models

huggingface/pytorch-image-models · ★ 37k ▲13/7d · Python

PyTorch Image Models, usually imported as `timm`, is a Python library that puts many image encoders, pretrained weights, data transforms, optimizers, and training scripts behind a shared PyTorch interface. It solves the tedious parts of comparing vision backbones and reusing their intermediate features without copying each paper's repository.

Verdict Our `timm` install consumed 5,432 MB and its test run exited 137 after 319 seconds at roughly 1% progress, even though the 7-second package build succeeded. Use it when a shared PyTorch API across many image backbones will save more engineering time than the large environment and model-by-model validation cost.

OpenMAIC

THU-MAIC/OpenMAIC · ★ 37k ▲2.8k/7d · TypeScript

OpenMAIC is an English-documented, open-source classroom generator that turns a topic or uploaded material into slides, quizzes, simulations, project work, and conversations with AI teachers and classmates. It can play lessons in the browser, edit generated scenes, and export presentations, interactive HTML, or video.

Verdict Our OpenMAIC build passed in 116 seconds and 4,892 of 4,954 tests passed, but installation consumed 2,672 MB and one guarded-import test failed. It is a strong trial candidate for educators who want to edit an AI-built lesson and can review every claim before delivery.

claude-plugins-official

anthropics/claude-plugins-official · ★ 36k ▲272/7d · Python

Claude Plugins Official is Anthropic's managed marketplace directory for Claude Code plugins. It contains Anthropic-maintained plugins and external entries from partners or the community, with optional commands, agents, skills, language servers, hooks, and MCP server connections.

Verdict Our Discord-plugin install took 46 seconds for 130 packages, yet the directory offered no build or test target for that measured path. Use Claude Plugins Official as a discovery index and convenient installer, not as a blanket trust decision.

DeepSeek-Reasonix

esengine/DeepSeek-Reasonix · ★ 36k ▲106/7d · Go

Reasonix is a Go-based coding agent that runs in a terminal, desktop app, browser, or editor through ACP. It gives long coding sessions a shared engine with permissions, workspace boundaries, checkpoints, model choice, and plugins, so developers can leave work running without surrendering every file and command to an opaque process.

Verdict Our 749-second test run ended with 25 of 26 Go packages passing, so Reasonix is buildable but not clean enough at commit b9cf32f for an unquestioned rollout. Use it if its shared CLI, desktop, web, and ACP engine solves a real workflow problem and you are prepared to pin a release.

diffusers

huggingface/diffusers · ★ 35k ▲64/7d · Python

Diffusers is a Python library for running and training diffusion models that generate images, video, audio, and other structured outputs. It packages pretrained pipelines, model components, and noise schedulers behind APIs that can be used as-is or rearranged for research and product code.

Verdict Our Diffusers install consumed 5,650 MB, pip-audit found 31 known vulnerabilities, and the test suite timed out at 900 seconds with failures already visible, so production adoption needs a narrowed dependency set and pipeline-specific tests. It remains the practical library choice for Python teams that need early access to many diffusion architectures and control over their parts.

Vibe-Trading

HKUDS/Vibe-Trading · ★ 34k ▲446/7d · Python

Vibe-Trading is a Python research workspace that connects an AI agent to market-data loaders, backtests, portfolio analysis, reports, broker journals, finance tools, and optional broker connectors. It helps turn a plain-language trading question into inspectable code and artifacts, while also exposing research tools through a CLI, web app, REST API, messaging channels, and MCP.

Verdict Our Vibe-Trading run passed 11,133 of 11,134 tests, with the only failure caused by a missing passwd entry for container UID 1000. That is strong evidence for its research code, and none for expected returns or safe live execution.

nuwa-skill

alchaincyf/nuwa-skill · ★ 33k ▲423/7d · Python

Nuwa's primary documentation and skill instructions are Chinese; a full English README is available. It directs an AI agent to research a public person's writing, interviews, decisions, critics, and speaking style, then turn that material into a reusable perspective skill with stated limits and sources.

Verdict Our Nuwa run installed 35 packages in 16 seconds and built in 5 seconds, but it had no test target, so the mechanics passed without validating a generated person's accuracy. Use it for an explicitly labeled perspective exercise when you will read the research notes and keep inference visible.

AIHawk

feder-cr/AIHawk · ★ 32k ▲1.3k/7d · Python

AIHawk is an open-source AI browser agent: you tell it in plain language to browse, click, type and read real web pages, and it drives an actual browser to do it. The problem it solves is slow, brittle, manual web workflows, like checking fares across calendar widgets, or filling forms that lack APIs.

Verdict AIHawk installed in 39 seconds, passed all 138 tests in 22 seconds, and exposed zero known vulnerabilities in our sandbox; if you can accept the OpenRouter key and the fact that a browser agent's judgment is only as good as the model you point it at, it is a legitimate shortcut for real web automation..

blender-mcp

MCPBlender/blender-mcp · ★ 29k ▲914/7d · Python

Blender MCP connects an MCP-capable AI client to a Blender add-on so the model can inspect a scene, edit objects and materials, run Python, and bring in external assets. It solves the awkward handoff between a text agent and Blender by translating MCP tool calls into commands over a local socket.

Verdict Our Blender MCP run installed 60 packages and passed all 20 tests in 35 seconds across install, build, and test steps. Use it for supervised scene prototyping where a technical artist can inspect each change and revert the Blender file.

blender-mcp

ahujasid/blender-mcp · ★ 29k ▲914/7d · Python

Blender MCP connects an MCP-capable AI client to a Blender add-on so the model can inspect a scene, edit objects and materials, run Python, and bring in external assets. It solves the awkward handoff between a text agent and Blender by translating MCP tool calls into commands over a local socket.

Verdict Our Blender MCP run installed 60 packages and passed all 20 tests in 69 seconds across install, build, and test steps. Use it for supervised scene prototyping where a technical artist can inspect each change and revert the Blender file.

mlflow

mlflow/mlflow · ★ 28k ▲100/7d · Python

MLflow is a server and SDK for recording machine-learning experiments, registering models, tracing agent calls, evaluating outputs, and controlling access to model providers. It gives teams one place to inspect runs and production AI behavior instead of assembling separate tracking, registry, and observability services.

Verdict Our MLflow run installed 116 packages and used 601 MB, then 273 tests passed before 200 collection or setup errors stopped the suite. Use it when experiment tracking, model governance, and agent observability genuinely belong on one operated platform.

shap

shap/shap · ★ 26k ▲31/7d · Jupyter Notebook

SHAP is a Python toolkit for explaining how model inputs contribute to individual predictions and broader model behavior. It gives data scientists a common set of explainers and plots for tree models, neural networks, text pipelines, and functions that do not expose their internals.

Verdict Our SHAP install took 91 seconds and its 4-second build passed, but test collection stopped in 5 seconds because matplotlib was missing. Use it when one attribution interface across several model families is worth learning the assumptions behind each explainer.

flash-attention

Dao-AILab/flash-attention · ★ 25k ▲45/7d · Python

FlashAttention is the official implementation of an exact attention algorithm that reduces data movement between GPU memory and faster on-chip memory. It supplies specialized training and inference kernels for supported NVIDIA and AMD GPUs, so model builders can replace a costly attention operation without changing its mathematical result.

Verdict Our FlashAttention checkout installed and built in 30 seconds, but pytest then stopped with 122 collection or setup errors and no executed test result. Use it when a supported GPU, an exact attention workload, and measured model-level gains justify owning a compiled extension.

t3code

pingdotgg/t3code · ★ 23k ▲561/7d · TypeScript

T3 Code is a control surface for coding agents already installed on your computer. Its web, desktop, and mobile clients connect to a local or remote server that owns Codex, Claude Code, Cursor, Grok Build, and OpenCode sessions, letting you manage threads, terminals, diffs, permissions, and agent work away from the original terminal.

Verdict Our T3 Code build succeeded in 60 seconds and 2,705 of 2,712 tests passed in 728 seconds with 0 failures, so this young control surface has unusually credible engineering evidence. Use it when remote phone access and one interface across Codex, Claude Code, Cursor, Grok, and OpenCode solve a daily problem.

watermarks-remover

guillaumemeyer/watermarks-remover · ★ 22k ▲566/7d · Python

watermarks-remover inspects and removes AI provenance signals and ordinary metadata from text, images, documents, web files, and audio or video containers that you own. It provides Python command-line tools, a local HTTP service, a Claude Code plugin, and optional research harnesses for checking several statistical watermark schemes.

Verdict Our watermarks-remover run installed 55 packages, built in 7 seconds, and passed its tests in 72 seconds with 0 known dependency vulnerabilities. That clean baseline makes it worth evaluating for authorized file hygiene, especially when outputs are written beside untouched originals.

peft

huggingface/peft · ★ 22k ▲41/7d · Python

PEFT is a Python library for adapting large pretrained AI models while training only a small fraction of their parameters. It tackles the cost and storage burden of full fine-tuning, and connects directly with Hugging Face Transformers, Diffusers, and Accelerate.

Verdict Our build finished in 13 seconds, but the test suite was still running when the 900-second cap stopped it. PEFT is the sensible default when you already use Hugging Face libraries and want adapter-based tuning without maintaining the method implementations yourself.

agency-agents-zh

jnMetaCode/agency-agents-zh · ★ 21k ▲146/7d · Shell

agency-agents-zh is a Chinese-language collection of 277 ready-made AI expert roles, with English documentation available through its linked upstream project. It gives teams reusable role instructions and workflows for coding assistants, including 64 roles tailored to Chinese platforms and industries, so they do not have to draft every specialist persona from scratch.

Verdict Our install finished in 4 seconds with 0 packages added. Use agency-agents-zh as a well-stocked Chinese role library, especially when its platform-specific experts save you from translating or localizing the English upstream yourself.

FunASR

modelscope/FunASR · ★ 20k ▲119/7d · Python

FunASR is a Python speech-recognition toolkit for offline transcription, streaming recognition, voice activity detection, punctuation, speaker diarization, emotion tags, training, and model deployment. It offers several checkpoint families instead of one fixed model, plus an OpenAI-compatible server and an MCP server for agent clients. The default README is English, with Simplified Chinese, Japanese, and Korean versions linked at the top.

Verdict Our FunASR run installed 94 packages and built in 13 seconds, but pytest executed 0 tests because collection ended with a `SystemExit: 2` error. Trial it when Chinese or multilingual speech, streaming, diarization, or edge deployment requires more choice than one Whisper model provides.

SkillSpector

NVIDIA/SkillSpector · ★ 17k ▲652/7d · Python

SkillSpector checks AI-agent skill files for prompt injection, data theft, dangerous code, excessive permissions, vulnerable dependencies, and other install risks. It reads the skill without running it, then combines static checks with optional language-model analysis to produce a score and recommendation.

Verdict Our run installed 103 npm packages in 35 seconds and used 232 MB, but the checkout exposed no build or test target, so it did not verify the Python scanner. Use SkillSpector as pre-install triage alongside coverage checks and human review, especially for skills with scripts or broad permissions.

claude-seo

AgriciDaniel/claude-seo · ★ 17k ▲323/7d · Python

Claude SEO is a Claude Code plugin that turns website auditing into a set of local commands and specialist agent workflows. It checks technical SEO, content, schema, search data, local visibility, and AI citation signals, then writes reports and prioritized actions instead of locking the findings in a vendor dashboard.

Verdict Our Claude SEO run installed 92 packages and built cleanly, but 2 tests failed when a dry-run hit an HTTP 403 rate limit. Use it as an open, inspectable audit assistant if Claude Code is already part of the job and a human will verify the recommendations.

xiaohongshu-mcp

xpzouying/xiaohongshu-mcp · ★ 16k ▲110/7d · Go

xiaohongshu-mcp is a Chinese-first MCP server that lets an AI client operate Xiaohongshu through an automated browser. A maintained English README exists. Its tools can log in, search and read posts, inspect profiles, publish images or video, and perform account actions such as commenting, liking, and saving.

Verdict Our xiaohongshu-mcp run installed 67 packages in 14 seconds and passed all 8 test packages, but that clean result cannot protect browser selectors from a Xiaohongshu page change. Use it for supervised local workflows where a person approves every public action and can repair login or search when the site moves.

speech-to-speech

huggingface/speech-to-speech · ★ 13k ▲69/7d · Python

Speech to Speech is a Python voice-agent pipeline that listens, turns speech into text, gets a language-model response, and speaks the answer. Each stage can be swapped, while applications connect through a documented subset of the OpenAI Realtime protocol over WebSocket or WebRTC.

Verdict Our speech-to-speech run produced 43 collection/setup errors and only 6 passing tests, even though installation and build both succeeded, so version 1.0.0 did not clear a basic source-checkout gate in our sandbox. It is still worth evaluating for experienced voice teams that need swappable local components and a Realtime-compatible client surface.

FluidVoice

altic-dev/FluidVoice · ★ 12k ▲196/7d · Swift

FluidVoice is a macOS dictation app that turns speech into text locally and inserts it into almost any app. It solves the usual privacy and subscription concerns of cloud dictation, while adding optional voice commands and AI cleanup.

Verdict FluidVoice is one of the strongest free Mac dictation trials because installation is simple, model choice is broad, and the core speech path can remain local. The private Fluid Intelligence runtime weakens the open-source pitch, and recent microphone and insertion bugs make it risky for work where a missed sentence is costly.

RapidRAW

CyberTimon/RapidRAW · ★ 10k ▲81/7d · TypeScript

RapidRAW is a desktop and experimental Android photo editor that leaves original images untouched and stores adjustments in sidecar files. It combines a RAW library, GPU color processing, masks, batch export, compositing tools, and optional generative editing for photographers who want an open alternative to subscription software.

Verdict Our RapidRAW run installed 427 packages, built in 5 seconds, and found 0 npm vulnerabilities, but the repository offered no test target. It is worth trying as a parallel editor for its sidecars, GPU workflow, local masks, and cross-platform packages.

pocket-tts

kyutai-labs/pocket-tts · ★ 9.5k ▲85/7d · Python

Pocket TTS is a local text-to-speech engine built to run on a CPU instead of depending on a hosted voice API or dedicated GPU. It gives Python applications, scripts, and a bundled local server streamed speech, supplied voices, and voice cloning across seven documented languages.

Verdict Our Pocket TTS install took 39 seconds and 41 of 42 tests passed, but the remaining 8-bit WAV test failed because `soundfile` was unavailable. That is a reasonable trial result for a developer who wants private CPU speech, provided audio validation and dependency checks are part of the application.

rf-detr

roboflow/rf-detr · ★ 9.5k ▲85/7d · Python

RF-DETR is a Python model family for finding objects, drawing instance masks, and, in preview form, locating human keypoints in images. It gives computer-vision teams one API for using pretrained models, fine-tuning them on COCO-format data, and exporting the result to several deployment runtimes.

Verdict Our RF-DETR install took 160 seconds and 5,778 MB, then its tests stopped on an unrecognized pytest option, so adopting the source tree requires more care than the short inference example suggests. It deserves a trial if detection and segmentation under one API fit your work and you can compare models on your own GPU.

Graft

NanoNets/Graft · ★ 8.1k ▲1.4k/7d · TypeScript

Graft builds a local graph of a codebase so coding agents can find systems, symbols, callers, and change impact without rereading the repository from scratch. Its structural layer uses tree-sitter, while an optional model-backed pass adds summaries and concept nodes for Claude Code, Codex, Cursor, Gemini, and other agent hosts.

Verdict Our Graft run built in 7 seconds and passed all 836 tests, but npm audit found 1 high-severity vulnerability, so the code earns a serious trial with a dependency review attached. Use it when coding agents repeatedly lose time mapping the same large repository and your team can inspect the generated context and configuration changes.

firecrawl-mcp-server

firecrawl/firecrawl-mcp-server · ★ 7.5k ▲35/7d · JavaScript

Firecrawl MCP Server lets an MCP-compatible agent search the web, scrape pages, crawl sites, parse files, and interact with pages through Firecrawl. It solves the adapter work between an AI client and Firecrawl's hosted or self-hosted API, with tool schemas, transports, authentication, retries, and rate handling already wired.

Verdict Our Firecrawl MCP Server run installed 341 packages, built in 5 seconds, and passed all 74 tests in 24 seconds, making the adapter easy to trial. Use it when Firecrawl is already your web-data service or when its hosted endpoint is an acceptable dependency.

koharu

koharu-rs/koharu · ★ 5.6k ▲40/7d · Rust

Koharu is a desktop workspace for translating manga while keeping the page artwork, text regions, cleanup, and typesetting in one project. It can detect text, run OCR, translate with local or hosted models, rebuild artwork behind the original lettering, and export finished pages as PNG or layered PSD files.

Verdict Our Koharu install consumed 2,172 MB and 543 seconds, then its 326-second build failed on a missing `atk` system library even though all 88 Vitest cases passed. Use a release build if you want a serious local-first manga workspace and have modern graphics hardware plus time to review every page.

gptme

gptme/gptme · ★ 4.4k ▲4/7d · Python

gptme is a local terminal agent that can edit files, run shell and Python commands, browse the web, inspect images, call MCP tools, and keep conversations across sessions. It supports several hosted model providers plus local models, and it can run interactively, from an editor through ACP, or as a scheduled headless agent.

Verdict Our gptme run passed 9,557 tests but still ended with 27 failures and 75 errors after 632 seconds, so the project is substantial and the current checkout needs environment-specific verification. It is a strong fit for an experienced terminal user who wants provider choice, local state, MCP, and deep customization.

router

workweave/router · ★ 4.4k ▲207/7d · Go

Weave Router is a source-available Go proxy that chooses an AI model for each request from a coding agent or application. It accepts Anthropic, OpenAI, and Gemini formats, scores each action locally, and sends it to an enabled provider so teams can use one endpoint instead of wiring every client to every model service.

Verdict Our run built Weave Router in 87 seconds and passed 106 of 106 tests, which makes the codebase worth a controlled trial. Use it when per-action model choice is worth operating Postgres, credential encryption, and another decision layer in the request path.

Unity-MCP

IvanMurzak/Unity-MCP · ★ 4.3k ▲49/7d · C#

Unity-MCP connects an MCP-capable coding agent to the Unity Editor and, with custom tools, to a compiled game. It lets the agent inspect scenes, edit scripts and assets, run tests, read profiler data, and call project-specific C# methods instead of merely suggesting code in chat.

Verdict Our Unity-MCP CLI run installed 64 packages, built in 12 seconds, and passed its tests in 33 seconds, making it an easy technical trial for a Unity team. Use it when you want an agent acting inside the editor and you will inspect every mutation in version control.

openresearch-cli

alphaXiv/openresearch-cli · ★ 3.3k ▲2.6k/7d · Rust

OpenResearch is a local workspace that turns Claude Code, Codex, or OpenCode into agents for literature review and code-based experiments. It keeps each experiment in a Git branch and worktree, records runs and evidence, and can send the same committed snapshot to local or remote compute.

Verdict Our OpenResearch run built in 92 seconds and all 1,338 tests passed, making it the most convincing source checkout in this group. Use it when agents will conduct real code experiments and Git-backed lineage is part of the research record; the fixed-command discipline is a useful guardrail, not paperwork.

sie

superlinked/sie · ★ 3.3k ▲36/7d · Python

Superlinked Inference Engine, or SIE, is a self-hosted server that puts embeddings, reranking, document conversion, structured extraction, safety checks, and open language models behind one API. It is meant to replace a pile of separate model services with a catalog that loads models as jobs need them.

Verdict Our SIE run installed 704 packages and used 1,597 MB, then passed its build and tests in 31 seconds combined, so the codebase is workable but hardly small. Use it when several private model services have become one platform problem and you already have people who can operate the cluster.

agent-of-empires

agent-of-empires/agent-of-empires · ★ 3.3k ▲47/7d · Rust

Agent of Empires is a session manager for running several coding agents from one terminal or web dashboard. It wraps installed tools such as Claude Code, Codex, OpenCode, Gemini CLI, and Oh My Pi in persistent tmux sessions, with optional worktrees and containers to keep parallel jobs apart.

Verdict Our Agent of Empires build succeeded, and 4,656 of 4,660 tests passed before four failures, including one tied directly to missing tmux, so the codebase is credible but the stated prerequisite is real. Use AoE when supervising multiple Claude Code, Codex, or OpenCode jobs has become harder than the jobs themselves.

harbor

av/harbor · ★ 3.2k ▲6/7d · Python

Harbor is a CLI and desktop companion for assembling local AI services with Docker Compose. It connects model backends, chat interfaces, search, speech, image generation, MCP tools, and coding agents so you can try a stack without wiring every container by hand.

Verdict Our Harbor checkout installed 97 packages in 19 seconds and used 121 MB, but it offered no build or test target, so the easy setup does not provide a full verification result. Use Harbor to explore and compare a local AI stack, especially when hand-wiring Compose services is slowing experimentation.

mjlab

mujocolab/mjlab · ★ 3.1k ▲62/7d · Python

mjlab is a Python framework for training robot behaviors across many MuJoCo simulations on NVIDIA GPUs. It brings Isaac Lab's manager-style environment building to MuJoCo Warp, so researchers can assemble scenes, observations, rewards, and commands without running Isaac Sim.

Verdict Our mjlab sandbox installed 149 packages using 5,765 MB, then its 135-second test step reported 1,067 passes and 0 failures. Use mjlab when your team has NVIDIA hardware, knows Isaac Lab's manager pattern, and wants MuJoCo Warp without the Isaac Sim platform.

MTPLX

youssofal/MTPLX · ★ 2.3k ▲155/7d · Python

MTPLX is a macOS application and Python command line for running local language models on Apple Silicon. It uses multi-token prediction heads already built into compatible models to draft several tokens, verify them together, and keep the normal sampling distribution when accepted.

Verdict Our MTPLX install took 138 seconds and its package built in 15 seconds, but tests failed after 69 seconds because the Debian environment had no `mlx`, so our run says nothing about its advertised decoding speed. Apple Silicon owners with a supported MTP model should trial its tuner against plain autoregressive decoding on their own Mac.

ainovel-cli

voocel/ainovel-cli · ★ 2.0k ▲43/7d · Go

The supplied documentation is primarily Chinese, and the provided material does not establish whether a full English guide exists. ainovel-cli is a Go command-line application that turns a short story brief into a long novel by coordinating planning, drafting, editing, and continuity tracking without requiring the author to drive every chapter.

Verdict Our build finished in 31 seconds, and all 52 reported tests passed, which makes ainovel-cli unusually credible for a fast-moving autonomous writing project. Use it if you want a structured, restartable novel factory and are comfortable reading Chinese documentation and paying for model calls.

webbrain

webbrain-one/webbrain · ★ 1.1k ▲62/7d · JavaScript

WebBrain is an open-source browser extension that lets an AI model read pages, answer questions, and carry out browser tasks in the tabs you already use. It can use its managed model, a cloud API, or a local OpenAI-compatible server, and it also lets coding agents delegate browser work through MCP.

Verdict Our WebBrain checkout installed 4 packages in 11 seconds and passed its available tests in 74 seconds, with 0 npm audit findings. Use it when an agent must work in the signed-in browser session you already have, especially through a local model or MCP client.

no-negative-echo

LB623/no-negative-echo · ★ 849 ▲66/7d · Python

No Negative Echo is primarily documented in Chinese, with a complete English README available. It is an Agent Skills-format instruction set and scanner that asks coding agents to write final titles, commits, pull requests, and handoffs from the accepted result instead of repeating ideas that were rejected during the conversation.

Verdict Our no-negative-echo run installed 35 packages and passed 105 collected tests in 15 seconds, making it cheap to inspect and try. Use it when rejected ideas routinely leak into commit messages or handoffs, especially after long agent sessions.

firecrawl

firecrawl/firecrawl · ★ 181k ▲2.5k/7d · TypeScript

Firecrawl is a web data API that can search, scrape, map, crawl, and interact with pages, then return markdown, structured JSON, screenshots, or page metadata. It gives agents and applications one interface for web work that may require JavaScript rendering or browser actions before extraction.

Verdict Our Firecrawl build failed in 9 seconds on an unresolved Rust-backed module, and 183 of 2,166 tests failed, so self-hosters should treat adoption as a platform project. The hosted API is the sensible first trial for teams that want search, crawling, extraction, and browser actions behind one interface.

firecrawl

mendableai/firecrawl · ★ 181k ▲2.8k/7d · TypeScript

Firecrawl is an API and self-hostable service that searches the web, turns pages into markdown or structured data, crawls sites, and controls browser sessions. It gives agents and applications one interface for pages that may need JavaScript rendering, link discovery, document parsing, or interaction before extraction.

Verdict Firecrawl is a convincing hosted choice when a product needs search, scraping, crawling, and browser interaction behind one API. The open code is useful too, but our unresolved Rust package, 183 failed tests, and multi-service development path make self-hosting a platform job rather than a cheap escape from API fees.

skills

anthropics/skills · ★ 177k ▲1.1k/7d · Python

Anthropic Skills is a collection of instruction folders, scripts, examples, and a specification that teach Claude how to perform repeatable tasks. It includes document creation skills, technical examples, a starter template, and plugin packages for Claude Code.

Verdict Anthropic Skills is a useful reference shelf, especially for seeing how a short instruction file can expand into scripts and supporting material. It is not a vetted package set to install wholesale.

ponytail

DietrichGebert/ponytail · ★ 139k ▲6.0k/7d · JavaScript

Ponytail is an English coding-agent skill that pushes an agent to solve a task with less new code: reuse the codebase, standard library, browser, platform, and installed dependencies before creating another abstraction. Spanish and Korean README translations exist. Plugins and rule-file adapters bring that guidance to Claude Code, Codex, Copilot CLI, OpenCode, Gemini, Hermes, and many other agent hosts.

Verdict Ponytail is a sharp corrective for agents that turn a date input into a component project. Its decision ladder is easy to understand, the adapters are unusually broad, and the repository stays small.

gstack

garrytan/gstack · ★ 133k ▲1.1k/7d · TypeScript

gstack is an opinionated collection of coding-agent skills that turns product planning, code review, browser QA, security checks, release work, and retrospectives into named commands. It began as a Claude Code setup and now installs profiles for Codex and several other agents.

Verdict Our gstack run installed 234 packages and built in 7 seconds, then the 264-second test step failed across 3 shards. The workflow is worth trying for a founder who wants strong opinions about planning, review, browser QA, and shipping, but the red suite makes a team-wide required install premature.

MoneyPrinterTurbo

harry0703/MoneyPrinterTurbo · ★ 124k ▲2.4k/7d · Python

MoneyPrinterTurbo is a Chinese-first Python application for generating short videos from a topic or script; maintained English and Japanese READMEs are available. It can write narration, find or generate footage, synthesize speech, add subtitles and music, render vertical or horizontal video, and optionally publish to social platforms.

Verdict MoneyPrinterTurbo covers an impressive amount of the short-video pipeline and passed our install, build, and test steps. The 28 known vulnerabilities prevent an easy production recommendation, especially with an API that stays open unless configured otherwise.

graphify

Graphify-Labs/graphify · ★ 118k ▲2.0k/7d · Python

Graphify turns a repository and its supporting documents into a local knowledge graph that an engineer or coding assistant can query. It parses code into named nodes and relationships, then produces an interactive HTML graph, a Markdown report, and JSON for path, explanation, and natural-language queries.

Verdict Our Graphify run passed 4,541 tests, but 24 Terraform cases failed because `tree_sitter_hcl` was unavailable, so the base tool looks well exercised while optional-language packaging still needs verification. Use it for faster orientation and scoped assistant context in a large repository, provided every consequential edge is checked against source.

awesome-design-md

VoltAgent/awesome-design-md · ★ 116k ▲1.2k/7d

awesome-design-md is a curated library of DESIGN.md files derived from the visual language of real technology websites. You copy one into a project so an AI coding or design agent has concrete guidance on typography, color, spacing, components, and overall feel instead of improvising from a vague style prompt. It is a reference collection, not a UI framework or executable application.

Verdict Our run produced no install, build, or test result because commit 8147538 has neither a supported language ecosystem nor a Dockerfile. Use awesome-design-md as a rich prompt library and learning resource, especially when an AI-generated page keeps drifting stylistically.

caveman

JuliusBrussee/caveman · ★ 106k ▲1.4k/7d · Go

Caveman is a set of local tools that tries to cut the text coding agents read and write. Its skill asks agents for shorter answers, while its proxy compresses request context and keeps recoverable copies of the original data on your machine.

Verdict Our Caveman install pulled 391 packages and used 915 MB, then its test command stopped after 4 seconds on a Node option the runtime rejected. Try the skill first if shorter agent answers are the goal; adopt the proxy only after checking output quality, credential routing, telemetry, and the BSL terms.

servers

modelcontextprotocol/servers · ★ 90k ▲200/7d · TypeScript

This repository contains seven small reference servers that demonstrate how Model Context Protocol clients can expose files, Git repositories, web fetching, memory, time, prompts, resources, and tools to a model. It is an educational code collection maintained by the MCP steering group, not the general MCP server directory or a production integration bundle.

Verdict Our build finished in 18 seconds, but no test target existed and npm audit found 5 known vulnerabilities, including 4 high-severity advisories. Use this repository to learn MCP shapes or test a client, exactly as its maintainers intend.

NextChat

ChatGPTNextWeb/NextChat · ★ 89k ▲52/7d · TypeScript

NextChat is a self-hosted chat interface for OpenAI, Anthropic, Google, DeepSeek, Azure, and compatible model APIs. It gives individuals or small teams one browser and desktop client for conversations, prompt templates, artifacts, plugins, and optional MCP tools while keeping chat data in the browser by default.

Verdict NextChat is a good fit for a personal or small-team model console when browser-local storage, provider choice, and quick deployment matter. The open-source edition is much weaker as an organizational service because its documented access control is a shared password, and current proxy security reports need an operator's attention.

taste-skill

leonxlnx/taste-skill · ★ 87k ▲1.9k/7d · JavaScript

Taste Skill is a collection of structured prompts, or "skills," designed to improve the design quality of AI-generated frontends. It solves the problem of AI agents producing generic, boilerplate user interfaces by giving them expert-level guidance on layout, typography, motion, and spacing.

Verdict Taste Skill is an essential tool for anyone serious about using AI for frontend development. It brilliantly addresses the most common weakness of AI-generated code, its lack of design taste, by packaging expert guidance into a simple, effective format.

worldmonitor

koala73/worldmonitor · ★ 86k ▲626/7d · TypeScript

World Monitor is a global intelligence dashboard that combines news, maps, markets, geopolitical signals, infrastructure data, and AI-generated briefs. It gives analysts and curious readers one situational view instead of making them watch many feeds and specialist sites separately.

Verdict Our World Monitor install consumed 2,346 MB, npm audit found 38 known vulnerabilities, and the repository exposed no test target, even though its 112-second production build passed. Use it as an exploratory intelligence desk when breadth, maps, and agent access outweigh a heavy operating surface.

Agent-Reach

Panniantong/Agent-Reach · ★ 82k ▲3.2k/7d · Python

Agent Reach's primary README is Chinese, and a substantial English translation is available in the repository. It gives command-running AI agents a maintained routing layer for reading and searching websites, social networks, video platforms, GitHub, and feeds through selected upstream tools.

Verdict Our Agent Reach run installed 45 packages, used 66 MB, and completed its test command successfully, so the core is cheap to trial. Use it when an agent needs several web and social sources and you accept that cookies, browser sessions, and upstream breakage remain your operational responsibility.

models

tensorflow/models · ★ 78k ▲3/7d · Python

TensorFlow Model Garden is Google's collection of TensorFlow model implementations, reusable training code, configs, and checkpoints. Its maintained `official/` package helps teams train or adapt vision, language, and recommendation models without rebuilding common architectures from papers.

Verdict Our TensorFlow Model Garden install took 78 seconds and occupied 2,894 MB, so it suits committed TensorFlow teams better than developers sampling one model. Use `official/` when you want Google's maintained training patterns and accept tight TensorFlow version alignment.

headroom

headroomlabs-ai/headroom · ★ 72k ▲1.5k/7d · Python

Headroom compresses tool output, logs, retrieved text, files, and conversation history before an AI model reads them. It can sit inside Python or TypeScript code, run as a local proxy, or expose compression and retrieval through MCP, with original content kept locally for later retrieval.

Verdict Our Headroom install pulled 218 packages and occupied 6,039 MB, while its test suite was still at 30% when our 900-second cap stopped it. Trial it if long agent sessions are expensive enough to justify a local proxy and you can compare compressed answers with the originals.

mem0

mem0ai/mem0 · ★ 65k ▲345/7d · Python

Mem0 is a memory layer for AI assistants and agents. It extracts facts from conversations, stores them through a chosen vector database, and retrieves relevant details so an application can remember a user across sessions.

Verdict Our Mem0 run installed 62 packages in 30 seconds and built in 7 seconds, but the suite ended with 36 failures and 117 collection or setup errors, so adoption should begin with backend-specific tests. It is a strong candidate for a team that wants one memory API across several agent stacks and accepts the work of choosing models, storage, and deployment.

daily_stock_analysis

ZhuLinsen/daily_stock_analysis · ★ 65k ▲265/7d · Python

Daily Stock Analysis is a Chinese-first project, with maintained English documentation, that collects market data and news, asks a language model to analyze selected stocks, and sends a daily decision dashboard. It covers several Asian and US markets through a web app, desktop client, command line, API, scheduled jobs, and chat-style analysis.

Verdict Our Daily Stock Analysis run installed 35 packages and built successfully, but its tests stopped after 8 seconds because `anyio` was missing, so commit e45d5ce is not ready for unattended use from the documented dependency set. It is worth trying for a technically capable investor who wants one multilingual watchlist report and will check each data source and conclusion.

strix

usestrix/strix · ★ 63k ▲1.2k/7d · Python

Strix is an open-source security testing agent that probes source code, web apps, and APIs, then records vulnerabilities with reproduction steps. It is meant to give developers a practical pentest loop without hiring a human tester for every code change, though it does not replace an independent assessment for high-risk systems.

Verdict Our Strix run passed 390 tests but failed 200 before pytest stopped, and pip-audit reported 3 known vulnerabilities. It is worth a controlled trial for security teams that understand pentesting and want an agent to widen routine coverage.

faceswap

deepfakes/faceswap · ★ 58k ▲11/7d · Python

Faceswap is a desktop and command-line application for replacing faces in images and video with a model you train on your own source material. It packages face extraction, model training, conversion, cleanup tools, and a GUI into one local workflow.

Verdict Faceswap remains a strong choice when you want to train and inspect the whole face-swapping pipeline, not merely upload a portrait and press a button. The GUI, detailed manuals, GPU options, and active codebase reward patient users, but the 71 setup errors in our test run stop us from calling the source checkout trouble-free.

ppt-master

hugohe3/ppt-master · ★ 55k ▲1.5k/7d · Python

PPT Master is a local skill and Python toolset that lets an AI coding agent turn documents, web pages, or a brief into an editable PowerPoint deck. It plans the argument, authors slide visuals as SVG, converts them into native PowerPoint objects, and can also fill existing templates or add charts, notes, transitions, animation, and narration.

Verdict PPT Master is one of the more serious open workflows for generating a deck that remains genuinely editable in PowerPoint. Its value comes from story planning, explicit design choices, and native object export, not from eliminating human review.

goose

block/goose · ★ 54k ▲283/7d · Rust

Goose is a local AI agent with a desktop app, command-line interface, and embeddable API for editing code, running tools, research, and repeatable workflows. The original `block/goose` address now resolves to `aaif-goose/goose`, where the project supports hosted and local model providers plus extensions through MCP.

Verdict Our Goose build and test commands each exceeded 900 seconds, so the packaged app is easier to recommend than source contribution on a 3-CPU machine. Use Goose when provider choice and MCP extensions are worth a powerful local agent, but approve its commands and inspect every shared recipe.

hyperframes

heygen-com/hyperframes · ★ 50k ▲2.0k/7d · TypeScript

HyperFrames turns HTML, CSS, media, and seekable browser animations into MP4 video through headless Chrome and FFmpeg. It supplies a CLI, rendering packages, reusable blocks, an editor, cloud paths, and agent skills so coding agents can plan and author videos in plain web files.

Verdict Our HyperFrames install consumed 1,641 MB, the build passed in 98 seconds, and the 317-second test command still exited 1 despite reporting 607 passing tests. Use it when plain HTML is the right interchange format between web developers and coding agents, and when your pipeline can inspect frames, duration, and audio before release.

upscayl

upscayl/upscayl · ★ 49k ▲196/7d · TypeScript

Upscayl is a desktop app that enlarges low-resolution images with Real-ESRGAN models running on your own computer. It gives Windows, macOS, and Linux users a graphical alternative to command-line upscalers, provided their machine has a Vulkan-compatible GPU.

Verdict Our Upscayl install consumed 1,418 MB and npm audit found 44 known vulnerabilities, even though its 51-second build succeeded. The packaged app is still an easy recommendation for a desktop user with a compatible GPU who wants local, visual image enlargement.

nanobot

HKUDS/nanobot · ★ 48k ▲295/7d · Python

Nanobot is a self-hosted personal AI agent that runs through a browser, terminal, API, or chat app. It gives one model access to files, shell commands, web tools, memory, MCP servers, scheduled jobs, and helper agents, so a developer can operate an assistant on their own machine or server.

Verdict Our sandbox installed Nanobot in 45 seconds and built it in 9, but 195 tests failed and pip-audit found 42 known vulnerabilities, so we would not deploy commit 3852956 without investigation. It is still a strong trial for experienced self-hosters who want one agent across a browser, terminal, API, and chat services.

codebase-memory-mcp

DeusData/codebase-memory-mcp · ★ 43k ▲718/7d · C

Codebase Memory MCP indexes source repositories into a local knowledge graph that coding agents can query through Model Context Protocol tools. It gives an agent structural search, call paths, change-impact maps, architecture summaries, code snippets, and cross-repository links without sending source code to a hosted indexing service.

Verdict Our graph UI build passed in 15 seconds and its tests passed in 10 seconds after a 352-package install, but we did not execute the native indexer those screens control. Codebase Memory MCP is worth piloting for local, multi-repository agent navigation when file-by-file exploration wastes context.

CodeWhale

Hmbown/CodeWhale · ★ 41k ▲51/7d · Rust

CodeWhale is a Rust coding agent that runs in a terminal, reads a repository, edits files, and executes commands with configurable approval rules. It supports hosted providers and local models, while keeping sessions, goals, agent roles, MCP connections, and recovery controls in one TUI.

Verdict Our CodeWhale install took 13 seconds and found 0 known npm vulnerabilities, but the package offered no build or test target, so the easy install is stronger evidence than the local verification story. Choose it if provider choice, resumable goals, MCP, and explicit access modes are central requirements.

VoxCPM

OpenBMB/VoxCPM · ★ 38k ▲721/7d · Python

VoxCPM is a local text-to-speech system that generates speech without first converting audio into discrete speech tokens. VoxCPM2 can read 30 languages, design a voice from a written description, clone a voice from reference audio, stream output, and support full or LoRA fine-tuning.

Verdict Our VoxCPM environment took 158 seconds and 6,394 MB to install, while pytest collected 0 tests and exited 3 despite a standalone check printing 22/22 passed. VoxCPM2 deserves a controlled trial when 30-language local speech, voice design, and cloning justify a large ML stack.

playwright-mcp

microsoft/playwright-mcp · ★ 37k ▲245/7d · TypeScript

Playwright MCP is a Model Context Protocol server that lets an AI agent browse and operate websites through Playwright. It sends structured accessibility snapshots to the model, giving the agent named elements and stable references for navigation, forms, screenshots, storage, network work, and test assertions.

Verdict Our Playwright MCP build took 6 seconds, but its tests stopped in another 6 seconds because the process used Node 18.20.8 and Playwright required Node 20 or newer. Use it when an MCP agent needs persistent, inspectable browser state and you accept the token and credential costs.

mediapipe

google-ai-edge/mediapipe · ★ 37k ▲86/7d · C++

MediaPipe is an open-source toolkit for putting machine-learning features into Android, iOS, web, desktop, edge, and IoT applications. It supplies cross-platform task APIs, ready-to-run models, customization tools, and a lower-level graph framework, so teams do not have to assemble every on-device vision, text, or audio pipeline from scratch.

Verdict Our install completed in 12 seconds, but this checkout offered no root build or test target. MediaPipe is a strong choice when you need Google's packaged on-device tasks across several platforms and accept a platform-specific integration path.

PageIndex

VectifyAI/PageIndex · ★ 36k ▲51/7d · Python

PageIndex builds a hierarchical tree from a document and lets a language model navigate that structure to answer questions, instead of splitting text into fixed chunks and searching a vector database. Its Python SDK can index and chat locally with your own model connection, or use a hosted service for OCR, image understanding, line-level citations, and corpus-scale features.

Verdict Our PageIndex run passed 345 tests with 0 failures in 181 seconds, while 50 tests were skipped and pip-audit found 1 vulnerability. Try it when long, structured PDFs defeat chunk-and-vector retrieval and you can pay for or host the model calls that tree search requires.

ai-website-cloner-template

JCodesMore/ai-website-cloner-template · ★ 34k ▲335/7d · JavaScript

AI Website Cloner Template is a prebuilt Next.js project plus an agent skill for recreating the visible front end of a website from one or more URLs. It tells an AI coding agent to inspect layouts, interactions, styles, content, and assets through browser automation, then build isolated routes and compare the result with screenshots.

Verdict Our copy installed 886 packages and built in 17 seconds, but it had no test target, so AI Website Cloner Template is a credible starting workflow rather than verified cloning software. Use it for authorized front-end reconstruction when a skilled developer will review the agent's assets, routes, responsive behavior, and legal scope.

VoiceStudio

debpalash/VoiceStudio · ★ 31k ▲9.1k/7d · Python

VoiceStudio is a local desktop and web workspace for cloning voices, generating speech, transcribing recordings, dubbing video, dictating text, and producing long-form audio. It brings several speech engines into one interface, so creators can keep core processing and project files on their own machines instead of paying for a hosted voice service.

Verdict VoiceStudio is a convincing choice for a technically comfortable creator who wants several private speech workflows in one local app. The successful install and build are encouraging, though the 849 MB dependency footprint, model downloads, platform rules, and current beta reports make this a machine to tend rather than an appliance to forget.

gbrain

garrytan/gbrain · ★ 30k ▲196/7d · TypeScript

GBrain is a self-hosted memory and knowledge layer for AI agents, built around Markdown files, a graph, search, synthesis, and background maintenance. It connects to Codex, Claude Code, OpenClaw, Hermes, and other MCP clients so an agent can recall information across sessions instead of relying on one chat history.

Verdict Our GBrain run built in 7 seconds and Bun counted 7,152 passing tests, but the full command still exited 1 after 869 seconds, so this commit did not clear its test gate. GBrain is compelling for a technical user who wants Git-backed agent memory and can own privacy, providers, and upgrades.

xiaozhi-esp32

78/xiaozhi-esp32 · ★ 30k ▲222/7d · C++

XiaoZhi ESP32 is firmware for building a voice assistant on ESP32 boards, with audio, displays, cameras, networking, and device control handled on the microcontroller. It connects the device to speech and language-model services, while MCP lets the assistant operate hardware such as speakers, LEDs, servos, and GPIO pins.

Verdict Our XiaoZhi ESP32 setup failed in 1 second because plain CMake could not load the ESP-IDF project file or recognize an ESP-IDF command. Use the prepared firmware if your board appears in the matrix and you want a quick route to a capable voice gadget.

frontend-slides

zarazhangrui/frontend-slides · ★ 29k ▲414/7d · JavaScript

Frontend Slides is a set of instructions and helper scripts that guides a coding agent to make self-contained HTML presentations from a brief or an existing PowerPoint file. It installs as a Claude Code plugin, while other agents with file and shell access can read the same skill directly. The result is editable web code rather than a deck made inside a conventional slide application.

Verdict Our Frontend Slides run installed 35 packages in 18 seconds and built in 11 seconds, while the checkout offered no test target. It is worth trying if you want an agent to design a one-file HTML deck and you will inspect the rendered slides, keyboard behavior, and exported files yourself.

page-agent

alibaba/page-agent · ★ 29k ▲65/7d · TypeScript

Page Agent is an English-documented TypeScript library that places a natural-language agent inside a web page; a Chinese README also exists. It reads a simplified text form of the DOM, asks a configured language model what to do, and can click or fill the interface without screenshots or a headless browser.

Verdict Our Page Agent build and tests passed, but npm audit reported 17 known vulnerabilities, including 3 critical and 9 high, so dependency review comes before embedding it in a signed-in application. Page Agent is a sharp choice for a copilot inside a web product you control because it works from the DOM and needs no remote browser.

agentmemory

rohitg00/agentmemory · ★ 28k ▲248/7d · TypeScript

agentmemory is a local memory service for coding agents, with MCP, hooks, skills, a viewer, and APIs for saving and recalling work across sessions. It supports Claude Code, Codex, Cursor, and other clients, while storing persistent state through a bundled iii engine.

Verdict Our agentmemory run passed build and tests, but its 299-package install used 1,235 MB and npm audit reported 14 known vulnerabilities, including 6 high-severity findings. Try it when cross-session coding context is painful enough to justify a resident service, then restrict capture and inspect recalled records before trusting them.

mlx

ml-explore/mlx · ★ 28k ▲73/7d · C++

MLX is an array and machine-learning framework designed around Apple silicon, with Python, C++, C, and Swift-facing APIs. It gives NumPy and PyTorch users familiar building blocks for training and inference, while its current packages also cover CPU-only and CUDA-backed Linux systems.

Verdict Our MLX build completed in 8 seconds, but pytest reached 0 test bodies and stopped with 33 collection/setup errors, so commit 2d27ab0 did not earn a clean Debian CPU baseline. MLX is still worth a focused trial for researchers on Apple silicon who want shared CPU and GPU memory with familiar Python APIs.

cs249r_book

harvard-edge/cs249r_book · ★ 28k ▲68/7d · Python

CS249r Book is an English-language, open-source curriculum about engineering machine-learning systems, with README translations in Chinese, Japanese, and Korean. Its two-volume textbook is joined by browser labs, a small teaching framework, infrastructure simulation, hardware exercises, interview practice, slides, and instructor material.

Verdict Our CS249r checkout was 1,003.3 MB before dependencies, and root npm exposed no build or test target, so cloning the whole curriculum is a poor first step for a reader. Use the online textbook first, then pull a specific lab, TinyTorch, or instructor path when you know which material you need.

mastra

mastra-ai/mastra · ★ 28k ▲266/7d · TypeScript

Mastra is a TypeScript framework for building agents, tool-using assistants, graph workflows, and other model-backed applications. It bundles model routing, memory, retrieval, human approval, evaluation, observability, deployment, and Model Context Protocol support so JavaScript teams do not have to assemble each layer separately.

Verdict Our Mastra test command was still running when it hit 900 seconds, after a successful 645-second build, so adopting the full monorepo requires serious CI time. Mastra is a strong fit for TypeScript teams that want agents, controlled workflows, storage, Studio, and MCP in one framework.

qwen-code

QwenLM/qwen-code · ★ 28k ▲157/7d · TypeScript

Qwen Code is an open-source coding agent that runs in a terminal, IDE, desktop shell, daemon, or programmatic SDK. It can use Qwen or other model providers to inspect a repository, edit files, run tools, coordinate subagents, and connect to MCP servers.

Verdict Our Qwen Code run installed 1,937 packages in 348 seconds, used 1,475 MB, and built successfully, but tests were still running at the 900-second cap and npm audit found 9 known vulnerabilities. Try it if provider choice, MCP, and several client surfaces outweigh the size and release pace.

lerobot

huggingface/lerobot · ★ 28k ▲172/7d · Python

LeRobot is a Python toolkit for collecting robot data, training control policies, and running those policies in simulation or on physical machines. It gives researchers one interface for supported arms, teleoperators, datasets, and models instead of making them wire each layer together themselves.

Verdict Our LeRobot install consumed 7,302 MB and its tests stopped after 19 seconds because UID 1000 had no username, so adopting it means owning a large Python environment before touching a robot. Use it when its shared dataset format, supported hardware, and policy catalog remove more integration work than they add.

dsh-desktop

anywhere-labs/dsh-desktop · ★ 27k ▲2.3k/7d · TypeScript

DSH Desktop is a Chinese-first desktop client that packages DeepSeek Harness for Windows and macOS; a substantial English README and English user and developer guides are available. It starts the local Harness service and adds a window, tray, terminal, updates, work profiles, and a plugin market without requiring users to install Node.js.

Verdict Our dsh-desktop run built in 22 seconds and passed 790 of 796 tests, with 1 lifecycle test timing out at 5,000 ms. The installers are a good choice for Windows or macOS users who specifically want DeepSeek Harness and its plugins without running the upstream command line.

baoyu-skills

JimLiu/baoyu-skills · ★ 26k ▲148/7d · TypeScript

Baoyu Skills is a collection of more than 20 agent skills for making images, diagrams, slides, formatted articles, translations, and social posts. It gives Claude Code, Codex, and other file-capable agents repeatable publishing workflows. The main README is English, and Chinese documentation is also available.

Verdict Our Baoyu Skills run passed 234 of 236 tests and reported 7 high-severity dependency findings, so it is useful publishing automation that still needs an install review. Choose a few skills for one defined workflow, inspect their instructions, and keep human approval before public posting.

vit-pytorch

lucidrains/vit-pytorch · ★ 26k ▲6/7d · Python

vit-pytorch is a collection of PyTorch implementations of Vision Transformer designs for image classification and related research. It lets researchers try many published architectures through compact Python APIs instead of rebuilding each paper from scratch.

Verdict Our build succeeded in 8 seconds, but `pip-audit` found 21 known vulnerabilities in the installed environment. Use vit-pytorch when you need an unusually broad, editable ViT research workbench and are prepared to own training and dependency review.

pentagi

vxcontrol/pentagi · ★ 24k ▲1.9k/7d · Go

PentAGI is a self-hosted system that lets AI agents research a target, run security tools in containers, keep findings, and produce penetration-test reports. It is built for authorized security work where a human wants to supervise or automate several stages from one web interface.

Verdict Our PentAGI run installed 670 packages and built successfully, but 4 of 117 tests failed, so this 400,161-line system needs a deliberate evaluation before it receives target access. Use it when you can isolate workers, control scope, budget model calls, and review the agent's commands.

index-tts

index-tts/index-tts · ★ 24k ▲154/7d · Python

IndexTTS is a local text-to-speech system that clones a voice from one reference recording and lets developers control emotion, pronunciation, and speaking speed. IndexTTS 2.5 covers Chinese, English, Japanese, Spanish, and Arabic through a Web UI, Python interface, or vLLM deployment recipe.

Verdict Our IndexTTS environment consumed 8,413 MB and pip-audit found 32 known vulnerabilities, even though the 9-second build and 52-second tests passed. Use it when 5-language voice cloning and emotion control justify a dedicated GPU speech stack.

MaxKB

1Panel-dev/MaxKB · ★ 23k ▲58/7d · Python

MaxKB is a web platform for building knowledge-base assistants and multi-step AI agents without writing the whole application from scratch. It combines document ingestion and RAG, a visual workflow builder, model connections, MCP tools, and embeddable chat, with English and Simplified Chinese README documentation.

Verdict Our MaxKB install pulled 224 packages, occupied 1,794 MB, and returned 7 known vulnerability findings, while the repository exposed no test target. Trial v2.10.5-lts when a visual RAG and agent platform can replace custom internal development, but change the default login immediately and validate every document type you depend on.

dyad

dyad-sh/dyad · ★ 22k ▲113/7d · TypeScript

Dyad is a desktop AI app builder that keeps the project and editing workflow on your Mac or Windows machine. You bring an AI provider key, describe the app you want, preview the result, and keep control of the generated files instead of working only inside a hosted builder.

Verdict Our Dyad run built in 143 seconds, but 9 of 6,350 tests failed and npm audit found 80 vulnerabilities, so this is a promising local builder with concrete maintenance work attached. Try it if owning the project files and choosing the model provider matter more than a spotless setup.

onnx

onnx/onnx · ★ 21k ▲60/7d · Python

ONNX is an open file format and specification for moving machine-learning models between frameworks, tools, and hardware. It gives teams a shared representation for model graphs, operators, and data types, reducing the need to tie training and inference to the same vendor stack.

Verdict Our installation did not finish before the 900-second cap. ONNX is still a strong default interchange contract when models must cross framework or hardware boundaries, backed by a current codebase and unusually detailed specifications.

skills

google/skills · ★ 20k ▲273/7d · Python

Google's `skills` repository is a collection of pre-built capabilities for AI agents, designed to automate complex tasks within the Google Cloud Platform. Instead of writing low-level API calls, this library provides high-level "skills", like deploying an AI model on Kubernetes or setting up a database, that an agent can use directly. It solves the problem of bridging the gap between a general-purpose AI agent and the specialized, intricate world of cloud management.

Verdict If your organization lives and breathes Google Cloud, `google/skills` is an essential toolkit for building the next generation of infrastructure automation. It packages Google's own expert knowledge into executable skills for AI agents, saving countless hours of development.

Speech

NVIDIA-NeMo/Speech · ★ 18k ▲48/7d · Python

NVIDIA NeMo Speech is a Python toolkit for building, adapting, and deploying speech systems, including speech recognition, text-to-speech, and speech-focused language models. It gives researchers and PyTorch teams one framework plus pretrained checkpoints, so they do not have to assemble every training and inference component themselves.

Verdict Our build succeeded in 9 seconds, but the tests were still at 17% when the 900-second cap ended, with nine export-related failures visible. NeMo Speech is a strong choice for an experienced NVIDIA GPU team that needs one serious framework across ASR, TTS, and speech LLM research.

nuclear

nukeop/nuclear · ★ 18k ▲43/7d · TypeScript

Nuclear is a free desktop music player that searches multiple sources, streams tracks, manages a library, and avoids ads or tracking. Its plugin system supplies music sources and metadata, while an optional MCP server lets AI clients control playback.

Verdict Our Nuclear build completed 8 of 9 tasks, then failed after 67 seconds because `cargo metadata` was unavailable, while all 5 Vitest tests passed. Download the packaged player if you want ad-free desktop streaming and are comfortable selecting source plugins.

mcp-for-beginners

microsoft/mcp-for-beginners · ★ 17k ▲48/7d · Jupyter Notebook

MCP for Beginners is Microsoft's free curriculum for learning the Model Context Protocol through explanations, security guidance, and sample servers and clients. It spans C#, Java, JavaScript, TypeScript, Python, and Rust, then branches into authentication, deployment, Azure integrations, databases, and newer protocol features.

Verdict Our `05-AdvancedTopics` run installed 64 packages and built in 8 seconds, but it exposed no test target, so the measured code path has no suite result. Use this curriculum for breadth, multilingual teaching, and security context, then verify protocol details against the official specification before shipping.

trigger.dev

triggerdotdev/trigger.dev · ★ 16k ▲55/7d · TypeScript

Trigger.dev runs long-lived TypeScript jobs and AI agent workflows with queues, retries, schedules, waits, logs, and deployment history. It gives application developers a place for work that outlives an HTTP request, either on Trigger.dev Cloud or on infrastructure they operate.

Verdict Our Trigger.dev checkout installed 2,536 packages and occupied 1,984 MB, then both build and test commands stopped at the database package, so this is a platform commitment rather than a small job runner. Use Trigger.dev Cloud when a TypeScript product needs observable, retryable AI work without owning the control plane.

text-to-cad

earthtojake/text-to-cad · ★ 16k ▲1.1k/7d · Python

text-to-cad is a set of coding-agent skills that turns written requests and reference images into local CAD, robot-description, drawing, slicing, and fabrication files. Its main CAD workflow writes editable Python, generates STEP geometry, checks dimensions and validity, and opens the result in a local viewer for review.

Verdict Our text-to-cad install used 1,561 MB for 87 packages and built in 6 seconds, but the measured CAD project exposed no test target. Use it when you want a coding agent to produce editable STEP-first work and you will follow its inspection and snapshot loop.

Figma-Context-MCP

GLips/Figma-Context-MCP · ★ 16k ▲60/7d · TypeScript

Framelink MCP for Figma is a Model Context Protocol server that turns Figma file and node data into a smaller description for coding agents. It helps an agent see layout, styling, and selected design structure without relying only on a screenshot.

Verdict Our Framelink MCP run installed 354 packages, built in 18 seconds, and passed its tests in 29 seconds. It is a focused choice for feeding selected Figma structure to Cursor, provided a developer owns the token and verifies every generated interface.

claudian

YishenTu/claudian · ★ 15k ▲91/7d · TypeScript

Claudian puts coding agents such as Claude Code, Codex, Grok, OpenCode, and Pi inside an Obsidian sidebar. It lets an agent read and edit vault files, search notes, run shell commands, use MCP tools, and carry out multi-step work without moving the conversation to a terminal.