agentgraph

agentgraph runs coding agents on your repo's issues. The agents, checks and workflows are Markdown files in .agentgraph/. Agents work in Docker sandboxes without a GitHub token, the engine pushes only checked work, and a person merges it.

Quickstart

You need Node 24, Docker without root and git config user.name. No API key: when you are logged in to Claude Code, the first agi start runs claude setup-token, tests the pasted token and keeps it in the macOS Keychain (a 0600 file on Linux). A ChatGPT login through Codex, an Anthropic key or LiteLLM also work; see Models.

In a repo with an off-by-one bug in padStart and a failing test:

npm install -g @jannik-stacks/agi
agi start "padStart is off by one when len is 0"
agi show 1
git diff main...agentgraph/1
agi review 1 changes "also cover a negative len"
agi merge 1
git merge --ff-only agentgraph/1
agi studio

The first run builds the sandbox images, about 3 minutes. An implementor changes the code and runs the tests, a reviewer reads it, then it waits for you. The branch lands in your repo; your checkout never moves.

Without .agentgraph/ it runs the default graph. No GitHub needed: ~/.agentgraph/<repo>-<hash>-local/ holds instance.yaml, the runs in data/ and forge/, which stands in for GitHub.

From source

agi studio needs a clone: run npm ci && npm run build && npm link in it, and npm run build again after each pull.

Docs

Start at What agentgraph is, then:

1. Install and first run
2. The graph folder
3. Checks and results
4. Workflows
5. Running on GitHub
6. Hosting
7. CLI

Examples

| Example | What it is | | --------------------- | ---------------------------------------------- | | examples/sandbox | A small JavaScript package; the place to start | | examples/agentgraph | The graph agentgraph runs on itself | | examples/bench | The graph the bench replays pull requests with |

Development

npm run check runs everything CI runs. See Engine and code and STYLE.md.

License

MIT, see LICENSE.