mrkeyoor.com_
Tue 01 Sept 17:40 UTC

Best Open-Source AI Coding Agents (2026) — measured, not ranked by stars

We cloned and ran every agent on this page in a clean container; the numbers are ours. Each one got the same treatment: a fresh checkout, the project’s own install command, then its own build and test commands, on a 3-CPU, 8 GB sandbox with a 900-second cap per step. Where a project exposes no build or test target, we say so instead of guessing. The order below is our usability score first, stars only as a tie-break — which is why a 27k-star agent outranks a 119k-star one.

Twelve terminal and editor agents, each cloned and run on our own machine — install time, dependency count, build and test results.

01

opencode

sst/opencode · ★ 203k ▲1.4k/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.

Measured sandbox lab-node:22 · 3 cpu / 8g · install 107s · 2,363 deps · 2,474 MB · no build target · tests FAILED after 15s · 681,820 LOC · 149s total

SetupDocsCommunityMaturity

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.

Pick it if Developers who want one coding-agent interface across many model providers, including local models.

Full review →
02

crush

charmbracelet/crush · ★ 28k ▲143/7d · 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.

Measured sandbox golang:1.24-bookworm · 3 cpu / 8g · install 117s · 452 deps · build 155s · tests 52/52 go test · 147,342 LOC · 477s total

SetupDocsCommunityMaturity

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.

Pick it if Terminal users who want one coding-agent interface across hosted and local models.

Full review →
03

codex

openai/codex · ★ 120k ▲2.4k/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.

Measured sandbox node:22-bookworm · 3 cpu / 8g · install 21s · 1 deps · 170 MB · no build target · no test target · 1,529,062 LOC · 48s total

SetupDocsCommunityMaturity

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.

Pick it if Developers who want an agent to inspect, edit, test, and explain a real local repository.

Full review →
04

gemini-cli

google-gemini/gemini-cli · ★ 107k ▲64/7d · 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.

Measured sandbox lab-node:22 · 3 cpu / 8g · install 47s · 1,325 deps · 795 MB · build 69s · tests 6,986/7,000 vitest · 680,799 LOC · 1,039s total

SetupDocsCommunityMaturity

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.

Pick it if Developers who already use Gemini and want an agent inside the terminal.

Full review →
05

OpenHands

All-Hands-AI/OpenHands · ★ 86k ▲684/7d · TypeScript

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

Measured sandbox node:22-bookworm · 3 cpu / 8g · install 227s · 1,998 deps · 1,043 MB · build 23s · tests passed in 476s (0 collected) · 279,276 LOC · 773s total

SetupDocsCommunityMaturity

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

Pick it if Engineering teams that want one interface for several coding agents and remote backends.

Full review →
06

cline

cline/cline · ★ 67k ▲404/7d · TypeScript

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

Measured sandbox lab-node:22 · 3 cpu / 8g · install 129s · 2,435 deps · 3,838 MB · build 64s · tests 72/72 vitest · 709,473 LOC · 736s total

SetupDocsCommunityMaturity

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

Pick it if Developers who want one coding agent across an IDE, terminal, scripts, and CI jobs.

Full review →
07

goose

block/goose · ★ 54k ▲247/7d · 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.

Measured sandbox lab-rust:1 · 3 cpu / 12g · install 46s · 1,281 deps · build FAILED after 900s (900s cap) · tests FAILED after 900s (900s cap) · 384,063 LOC · 1,871s total

SetupDocsCommunityMaturity

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.

Pick it if Developers who want an agent to inspect files, edit code, and run project commands from a desktop or terminal.

Full review →
08

qwen-code

QwenLM/qwen-code · ★ 28k ▲162/7d · 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.

Measured sandbox lab-node:22 · 3 cpu / 8g · install 348s · 1,937 deps · 1,475 MB · build 236s · tests FAILED after 900s (900s cap) · 3,302,842 LOC · 1,514s total

SetupDocsCommunityMaturity

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.

Pick it if Developers who want a terminal agent while retaining a choice of Qwen, OpenAI, Anthropic, Gemini, or local model endpoints.

Full review →
09

aider

Aider-AI/aider · ★ 49k ▲143/7d · Python

Aider is a terminal coding assistant that sends repository context to an LLM, applies edits, and works with Git so developers can inspect or undo what changed. It supports hosted and local models, codebase maps, lint and test commands, images, web pages, and an editor watch mode.

Measured sandbox ghcr.io/astral-sh/uv:python3.12-bookworm · 3 cpu / 8g · install 104s · 171 deps · 772 MB · build 9s · tests 473/492 pytest · 42,721 LOC · 790s total

SetupDocsCommunityMaturity

Verdict Our Aider run passed 473 tests, but 10 failed, 9 ended in setup errors, and pip-audit reported 126 known vulnerabilities after a 772 MB install. It remains a strong fit for an experienced developer who wants model choice and Git-visible edits in the terminal.

Pick it if Developers who prefer a terminal and Git diff over a separate AI editor.

Full review →
10

kilocode

Kilo-Org/kilocode · ★ 27k · TypeScript

Kilo Code is an open-source AI coding agent that works in VS Code, JetBrains IDEs, the terminal, and a web-hosted cloud agent. It handles code generation, repository questions, debugging, review, autocomplete, and command execution, so developers do not need a different assistant for each interface.

Measured sandbox lab-node:22 · 3 cpu / 8g · install 80s · 1,964 deps · 2,609 MB · no build target · tests FAILED after 18s · 1,157,861 LOC · 141s total

SetupDocsCommunityMaturity

Verdict Our install completed in 80 seconds, but it pulled 1,964 packages and the root test command failed by design after 18 seconds. Kilo Code is worth trying if you genuinely want the same agent across VS Code, JetBrains, CLI, cloud, and automated review, especially when model choice matters.

Pick it if Developers who want one coding agent across VS Code, JetBrains, and the CLI

Full review →
11

continue

continuedev/continue · ★ 36k ▲78/7d · TypeScript

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

Measured sandbox node:22-bookworm · 3 cpu / 8g · install 16s · 319 deps · 107 MB · no build target · no test target · 331,015 LOC · 49s total

SetupDocsCommunityMaturity

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

Pick it if Developers who want an open codebase for a coding agent across the terminal and two major IDE families.

Full review →
12

plandex

plandex-ai/plandex · ★ 16k ▲3/7d · Go

Plandex is a terminal coding agent built for changes that span many files and many model calls. It keeps generated edits in a reviewable sandbox, manages large code context, runs commands, and can work with several model providers before you apply anything to the project.

Measured sandbox golang:1.24-bookworm · 3 cpu / 8g · install 49s · 360 deps · build 40s · tests passed in 13s (0 collected) · 81,449 LOC · 115s total

SetupDocsCommunityMaturity

Verdict Our Plandex run installed 360 packages and built in 40 seconds, but its successful test step found 0 tests and the hosted service no longer accepts users. The cumulative diff and large-context workflow still merit a trial for an experienced self-hoster.

Pick it if Developers handling large repository changes who want a plan, cumulative diff, and explicit apply step in the terminal.

Full review →

Also run, not ranked: Roo-Code installed 2,665 packages and passed build and tests, but the extension was shut down and the repository archived on 15 May 2026, so it is not a 2026 pick.

See also