Autonomous AI Agent · Written in Rust

Your autonomous AI software engineer

Forge takes a GitHub issue, spins up an isolated Docker sandbox, autonomously writes and tests code, and produces a verified git diff ready to merge.

Get Started View on GitHub

scroll

What Forge does

A complete autonomous engineering pipeline, from issue to branch.

Docker Sandbox

Every run is fully isolated — no leftover state, no host pollution.

Any OpenAI-compatible Model

Works with OpenAI, Gemini, Anthropic, Ollama, or any compatible endpoint.

Autonomous Agent Loop

Thinks, acts, observes, repeats until the task is done or the step limit is reached.

Full Trajectory Recording

Every step, command, and model response saved to a .traj file.

ElizaOS Integration

Drop-in action plugin for ElizaOS AI agents.

Auto-fix on Label

Label any issue 'forge' — Forge picks it up, fixes it, and pushes branch forge/issue-{N} automatically. No commands to run.

How it works

Eight autonomous steps from issue to pull request, no human in the loop.

forge — bash

1Fetch the problem statement (GitHub issue, text, or file)
2Start an isolated Docker sandbox
3Clone the repository
4Enter the agent loop — think, act, observe
5Execute bash commands autonomously
6Run submit to capture the git diff
7Save the full trajectory to a .traj file
8Push branch forge/issue-{N} — review and merge when ready

Get started

You pick the issue. Forge does the work.

No Rust · No cloning · No compiling · Just Docker + .env

Install Docker

That's the only prerequisite. No Rust, no compiling, no cloning.

bash

docker --version

Create a .env file

Pick your model provider, paste your API key, and optionally add a GitHub token for automatic PR creation.

.env

# .env — create this file anywhere and run from that directory

# Google Gemini (recommended)
FORGE_MODEL=models/gemini-2.0-flash-001
FORGE_BASE_URL=https://generativelanguage.googleapis.com/v1beta/openai
FORGE_API_KEY=your-gemini-api-key

# OpenAI
# FORGE_MODEL=gpt-4o
# FORGE_BASE_URL=https://api.openai.com/v1
# FORGE_API_KEY=sk-...

# GitHub token — enables automatic pull request creation after each fix
# Create one at: github.com/settings/tokens  (needs repo scope)
GITHUB_TOKEN=ghp_...

# Find your Docker group GID: getent group docker | cut -d: -f3
DOCKER_GID=132

With GITHUB_TOKEN set, Forge automatically opens a pull request after every successful fix.

Run against a GitHub issue

Pass the repo and issue number. Forge pulls the image, clones the repo inside a sandbox, works autonomously, and opens a PR when done.

bash

docker compose run --rm \
  -e FORGE_REPO=owner/repo \
  -e FORGE_ISSUE=42 \
  akachiokey/forge:latest

Forge clones the repo internally — you never need to clone it yourself.

Enable always-on watch mode

Start once, fix forever. Add your repo and token, start the watcher, then just label issues on GitHub — Forge does the rest.

bash

# Add to .env:
FORGE_WATCH_REPO=owner/repo
FORGE_WATCH_LABEL=forge
FORGE_WATCH_INTERVAL=60
GITHUB_TOKEN=ghp_...

# Start in the background:
docker compose up watch -d

On GitHub, add the label 'forge' to any issue. Within 60 seconds Forge picks it up, fixes it, and pushes a branch forge/issue-{N} for you to review and merge.

Any OpenAI-compatible model

Forge is model-agnostic. Three lines in your .env is all it takes.

.env

FORGE_MODEL=gpt-4o
FORGE_BASE_URL=https://api.openai.com/v1
FORGE_API_KEY=sk-...

Copy the relevant block into your .env file. These three variables are the only ones required to start.

All environment variables

Variable	Required	Description
FORGE_MODEL	Yes	Model identifier passed to the API
FORGE_BASE_URL	Yes	Base URL of an OpenAI-compatible completions endpoint
FORGE_API_KEY	Yes	API key for the model endpoint
FORGE_REPO	No	GitHub repo for one-shot mode (owner/repo)
FORGE_ISSUE	No	Issue number for one-shot mode
FORGE_WATCH_REPO	No	GitHub repo to monitor in watch mode
FORGE_WATCH_LABEL	No	Label to watch for (default: "forge")
FORGE_WATCH_INTERVAL	No	Seconds between polls (default: 60)
GITHUB_TOKEN	No	GitHub PAT — raises rate limit; required for private repos