A local-first, Claude-powered engineering companion: specialist agents (hobits) that watch your codebases, run analyses, and surface only what matters.

Tech leads, chief and principal engineers, CTOs and VPs who own many codebases at once and want continuous visibility without alert fatigue.

A React UI talks to a FastAPI backend on PostgreSQL with pgvector and TimescaleDB. Deterministic scanners run continuously; a Postgres-backed job engine runs the Claude Code CLI headless to reason over what changed, and Prefect handles scheduled hobit runs. Every Claude call is a tracked job with token and cost accounting.

On a Claude Pro/Max subscription, hobits run through the Claude Code CLI at no extra cost; an API key works too. Either way, deterministic scanners do the heavy lifting and scheduled runs skip repos that haven't changed, so token spend tracks actual change — not the size of your fleet.

Yes. Your repos and data stay on your machine, and the API is meant for localhost or a trusted network. Claude is the only external call, and it sees only what changed.

Yes. Rate any answer with a thumbs, score or note; it becomes a lesson the hobit folds into its next runs on its own. Lessons are plain text you can read and edit, and each hobit's autonomy level is a setting you control.

Only when you ask. Click “Implement with AI” on a roadmap ticket and Shire drives the work through to a merge request you review. It never touches your code unprompted.

Clone the repo and run ./setup.sh. Docker brings up the UI on localhost:3000 and the API on :8000; a native dev path is documented too. You'll need a Claude subscription token or an API key.

Honestly: it's young. It started as a prototype, shipped real value fast, and is now MIT-licensed and moving quickly. Expect rough edges; contributions are welcome.