Autonomous coding agents, in parallel one trunk, many branches
Run agents on a project until all features pass.
Parallel agents. Feature DAG. Kanban UI. Provider pool.
Why claw-forge
Built for teams that outgrow a single API key and need real production reliability.
Never hit rate limits again. Anthropic, Bedrock, Azure AI Foundry, Vertex AI, Groq — auto-fallback with circuit breaker and RPM tracking per provider.
6 LSP servers (Python, Go, Rust, TS, Solidity, C++) + systematic debugging + verification gate + parallel dispatch — ready out of the box, no config required.
No subprocess+threading mess. TaskGroup + Semaphore. Clean, fast, and debuggable. Waves of parallel tasks with dependency ordering.
Add new agent types via pyproject.toml entry points. Third-party plugins just work — no patching the core, no forks.
Real-time board with WebSocket updates. Provider health dots. Cost tracker. Needs-Human column for stuck agents. No build step — just open your browser.
Use your claude login credentials. Zero API key juggling for the primary provider. Add keys for extra redundancy, not as a requirement.
Architecture
The dispatcher routes tasks through the provider pool, agents run in parallel waves, and everything streams live to the Kanban UI.
The Kanban board tracks all features across Pending · Running · Passing · Failed · Blocked columns in real time
Install with uv for instant, isolated installation. No virtualenv ceremony.
pip install uv
uv tool install claw-forge
claw-forge init my-project
claw-forge run my-project
# YOLO mode: max speed
claw-forge run my-project --yoloAn open-source harness that runs multiple AI coding agents in parallel against a single repository. You give it a written specification; it decomposes that into a dependency graph of features, dispatches an agent per feature into its own git worktree, and merges each one only after your own test suite passes.
A single agent works one task at a time in your working tree. claw-forge schedules many concurrently, each in an isolated worktree, so they cannot overwrite one another. It also tracks dependencies between features, so a task only starts once the work it builds on has merged.
Eight provider types: Anthropic API, Anthropic OAuth, Anthropic-compatible endpoints, Amazon Bedrock, Google Vertex AI, Azure, OpenAI-compatible endpoints, and local Ollama. Requests route through a pool with per-provider circuit breaking, so a rate-limited or failing provider is skipped rather than failing the run.
Each agent works in its own git worktree, never your checkout. File and shell tool calls are confined to that worktree, and on macOS and Linux an OS-level filesystem sandbox denies writes to the parent project. Work merges only after the acceptance gate runs your real test command and it exits zero.
Yes — Apache-2.0. You supply your own provider credentials, so the only cost is whatever those providers charge for the tokens your agents consume.
Python 3.11 or newer, a git repository, and credentials for at least
one model provider. Install with
uv tool install claw-forge.