DevClaw is an open-source autonomous AI coding kernel written in Go. One binary, zero deps, any OpenAI-compatible LLM. Five roles, thirteen built-in tools, parallel sub-drones, git worktree isolation.
or from source: go install github.com/yinhe/devclaw/cmd/drone@latest
Pick your OS. The binary is ~2.5 MiB and has zero runtime dependencies.
# 1. Download the release archive for your platform from GitHub # (replace ARCH with x86_64 or arm64, OS with linux or mac) curl -fSL "https://github.com/yinhe/devclaw/releases/latest/download/drone_0.1.2_linux_x86_64.tar.gz" \ | tar xz # 2. Move to PATH sudo mv drone /usr/local/bin/ # 3. Smoke test drone version drone roles # 4. Run against local Ollama (free, offline) or any OpenAI-compatible API ollama pull qwen3-coder drone run "add JSDoc comments to all exported functions in src/utils.js"
# Download + extract Invoke-WebRequest -Uri "https://github.com/yinhe/devclaw/releases/latest/download/drone_0.1.2_windows_x86_64.zip" -OutFile drone.zip Expand-Archive drone.zip -DestinationPath . # Smoke test .\drone.exe version .\drone.exe roles
go install github.com/yinhe/devclaw/cmd/drone@latest
DevClaw is a kernel. Everything here is open-source under Apache-2.0. Nothing below requires a paid plan.
Everything you need to run autonomous coding on your own machine.
bash, file_read/write, multi_edit, agent, parallel, undo, bash_approval, …dev, test, ops, sense, scout × 3 permission tiersThe StarClaw team maintains closed-source tooling that wraps this kernel for internal use.
Forge — issue tracking & orchestration surfacePheromone — event bus & task poolOverlord — fleet management for many dronesAbathur — trajectory → distilled skills engineBuilt to be useful on your laptop today, not to upsell you tomorrow.
Agent and Parallel are built-in tools. One drone can spawn up to 5 sub-drones and collate their work. A task tree, not a loop.
Five roles × three permission tiers (readonly, workspace_write, full_access) enforced in the runtime, not the prompt. Safe by default.
DRONE.md and .drone/skills/*.md are auto-injected into every run — your conventions, gotchas, and patterns travel with the drone.
First-class MCP stdio client. Drop any community server into .drone/mcp.json and the new tools respect the same permission tiers.
Every task gets its own git worktree. Drones never clobber each other's work — parallelism without merge hell.
Pre-built releases for linux/mac/windows × amd64/arm64. ~2.5 MiB. Runs behind any firewall with any OpenAI-compatible LLM.