Status
gitgoblin is in early development, built in the open. Discovery and repository analysis work from a clone today and from any MCP client through the server released on GitHub. The whole flow runs through one command, gitgoblin: discover, analyze, contribute, then publish once you have read the change. Preparing a contribution is experimental and costs money per run. Nothing is published to npm. This page is kept honest by hand; when it disagrees with the README’s status table, the README wins.
Last updated 2026-10-23.
What exists
| Part | Package | Status |
|---|---|---|
| Shared types: repository identity, discovery goals and signals | packages/core |
Working |
| Clients for GitHub and package registries, with a local cache | packages/sources |
Working |
| Ranking repositories for a goal, with reasons | packages/discovery |
Working |
The gitgoblin command: discover, explain, goals, cache, analyze, contribute, runs, show, publish |
packages/cli |
Working from a clone; not published to npm |
Repository profiling and opportunity discovery (analyze) |
packages/intel |
Working from a clone; scores still being calibrated |
| Plan, implement, review, verify, prepare a pull request | packages/engine |
Runs end to end on a test repository; experimental. See its README |
| MCP server for Claude Code, Codex, Cursor, Gemini CLI, VS Code and other agents, plus an agent skill | packages/mcp |
Released on GitHub (mcp-v0.1.0) and listed in the MCP Registry; npm publish pending. See docs/agents.md |
| Brand, documentation and this website | brand/, docs/, site/ |
Live |
Being built now
- Calibrating the scores behind
analyze, so rankings across repositories mean the same thing. - Precision of
analyze: issues that are still open but already fixed on the default branch, and files that merely quote a deprecated API, still show up as opportunities. The engine catches these during its first phase, but that costs a run. - The engine’s install step for Python projects, which needs a per-run virtual environment to work inside the sandbox.
- The maintainer opt-out list in docs/excluded-projects.md, enforced in discovery and the engine rather than only by people.
- Real examples. The walkthrough on the front page is scripted from the design and the engine’s command line; a recording of a real run will replace it.
Deliberately later
- Running many tasks in parallel, or on a schedule.
- Container or VM isolation for the engine. Version 0.1 uses the Agent SDK sandbox plus hooks; the interface allows stronger isolation later.
- Reacting to CI results or maintainer review after a pull request is open.
If you want to help with any of this, start with CONTRIBUTING.md.