@agile-graph/engine

An AI agile team for your repository. You are the product owner: you write feature specs with acceptance criteria, approve plans and accept results. AI workers take the team's roles (developer, tester and reviewer), and every change must pass the checks, code review and security scan you configure before it is offered to you as done.

The team works locally, in small, checked increments. Project context and reviewed memory stay on your machine, connected AI clients retrieve source-backed context through MCP, and each step runs in an isolated Git worktree, limited to the files it declares.

Install

npm install --global @agile-graph/engine
graph-engine --version

You need Node.js 22 (22.18 or later) or 24 to 26 and Git. Docker is needed for managed verification and security scans, not for indexing or retrieval.

Start

# Cloud AI client (Claude Code, Codex, Cursor):
graph-engine -C /path/to/your/project init --cloud-mcp --export 'src/**'
# Local model only:
graph-engine -C /path/to/your/project init
graph-engine -C /path/to/your/project index --lexical
graph-engine -C /path/to/your/project serve

serve prints a loopback URL for the dashboard. A new project uses local-only inference, no external network, no publication and shadow-mode decisions. No model, credential or provider is set up for you: add the ones you choose with graph-engine provider-add, then plan work with graph-engine plan, approve it with graph-engine plan-approve and accept results with graph-engine accept. graph-engine --help lists every command.

To give an AI client such as Claude Code, Codex or Cursor the project's context, run graph-engine -C /path/to/your/project mcp as its MCP server. See connecting an AI client.

Optional integrations

These are optional peer dependencies. Install one next to the engine only if you want it:

• Local embeddings: npm install --global @huggingface/transformers. On Linux x64 its install also downloads GPU runtime binaries; set ONNXRUNTIME_NODE_INSTALL=skip to skip them and run on the CPU.
• Cursor workers: npm install --global @cursor/sdk. Cursor's SDK has its own proprietary license, and it brings an older HTTP client with known advisories, so install it only if you use Cursor as a worker.

The security gate runs its scanners in a local Docker image. Build it once from the Dockerfile this package ships:

docker build -t graph-security:local "$(npm root --global)/@agile-graph/engine/scanner"

Data and privacy

Each project's databases, runs and workspaces are stored in your per-user data directory (~/Library/Application Support/graph-engineering on macOS, ~/.local/share/graph-engineering on Linux, %LOCALAPPDATA%\graph-engineering\Data on Windows), never in the package. Nothing leaves your machine unless a project policy you write allows it.

Automatic difficulty reports are on by default but only send when a project's policy allows network access to api.github.com. They carry fixed error kinds and platform metadata, never source, paths, prompts or logs, and are filed as public GitHub issues under your own GitHub CLI login. Turn them off with graph-engine feedback-config off or GRAPH_ENGINE_NO_FEEDBACK=1.

Documentation

The repository has the platform guide, working as an AI agile team and the documentation index.

Security

Report vulnerabilities privately. See the security policy.

License

MIT © 2026 Milton Adina. Third-party material and its licenses are listed in NOTICE.