github.com/ForgeAILab/forge @v0.3.0

Make multiple coding agents collaborate on one repo — without stepping on each other.

Forge gives every task its own isolated git worktree, then runs your CI gate and review step before changes touch main. Drive Claude Code, Codex, Cursor, Gemini, opencode, and your own shell agents through one local-first task lifecycle — REST API, MCP endpoint, CLI, and web UI in a single binary. Self-hosted, MIT, no cloud.

Quickstart · Why Forge · Docs · Changelog · Contributing

Why Forge

Running two coding agents against the same repo is how you lose diffs. Forge fixes that: every task runs in its own git worktree, hits your CI gate, and waits for review before it merges. Agents collaborate; they don't collide.

• One isolated git worktree per task — Claude Code, Codex, Cursor, and Gemini can each work in parallel without overwriting each other or polluting your main checkout.
• Review gate with your CI — define ci_steps per task; the review runner blocks merge until they pass. Human approval is an explicit transition, not an afterthought.
• Structured task lifecycle — todo → in_progress → review → merging → done, with an audit log and explicit cancellation paths so handoffs between agents (and humans) are legible.
• BYO agent — first-class adapters for Claude Code, Codex, Cursor, Gemini, opencode, and a generic shell executor. Add your own with a small adapter.
• One binary, every surface — REST API, MCP JSON-RPC, forge-ctl CLI, and a React web UI ship together. Drive Forge from a script, an editor, or a browser.
• Local-first by default — single binary, SQLite, loopback-only server with a persisted local port. No telemetry, no SaaS, your data stays on disk.

Who it's for

• Developers already running more than one coding agent (Claude Code + Codex, Cursor + Gemini, …) who keep losing diffs to branch collisions or shared-tree edits.
• Small engineering teams piloting agent workflows who need worktree isolation, audit trails, and a review gate before code lands on main.
• Builders who want a local, hackable control plane for AI coding work — not another hosted dashboard.

If you only want a chat UI bolted onto your editor, Forge is not for you. Try Continue or Cline instead.

5-minute quickstart

# Run instantly through npm (macOS / Linux)
npx @forgeailab/forge --demo

# Install via Homebrew (macOS / Linux)
brew install forgeailab/tap/forge

# Or grab the latest release directly
curl -fsSL https://raw.githubusercontent.com/ForgeAILab/forge/main/install.sh | bash

# Start the server with seeded demo data if you installed it locally
forge --demo

Open the management_url printed in the server logs. That's it — you should see a demo project with a labelled task and a fake daemon report. From here:

• Drive a real task end-to-end → docs/getting-started.md
• Wire up Claude Code / Codex / Cursor / Gemini → docs/getting-started.md#agents
• Hit the API directly → docs/api.md

Prefer to build from source? cargo run -p forge-cli -- --demo.

Core concepts

Task detail — lifecycle, role assignments, CI/review gate, audit log	Daemon — auto-detects installed CLIs on the host (Claude Code, Codex, Cursor, Gemini, opencode, shell)

Deeper dive → docs/architecture.md.

Documentation

Status

Forge is in public beta (0.1.x). The local-first single-user product is usable end-to-end, but APIs, schemas, and CLI flags can change without deprecation cycles. Track breaking changes in CHANGELOG.md. A stable 1.0 will land once the workflow engine, multi-user story, and release artifacts (signing, SBOMs, Homebrew, Windows builds) are finalized.

Contributing

Issues, PRs, and design discussion are welcome. Start with CONTRIBUTING.md and check good first issue and help wanted labels. By participating you agree to the Code of Conduct.

Security

Please report vulnerabilities privately per .github/SECURITY.md.

License

MIT © Forge contributors.