An open-source AI lab shipping tools, agents, and experiments across the AI development stack.
Reusable skills and harnesses that make coding agents sharper over time.
- autoharness (665 stars) — A self-learning skill layer for Claude Code: distills skills from your real sessions, updates them as you work, and prunes the ones that go unused.
- design-harness (113 stars) — Feed your agent papers and half-formed ideas; it links them into a system design you can defend. An Agent Skill for Claude Code and any SKILL.md-compatible agent.
See and control what your agents actually send — cost, context, and data boundaries.
- cost-xray (299 stars) — See what Claude Code and Codex actually send to the API — and what each part costs.
- phi-context-boundary-report — PHI boundary reports for healthcare and insurance AI workflows — across prompts, RAG, tools, memory, and logs.
Turn real engineering work into distribution, without the AI-slop.
- auto-gtm (25 stars) — A Claude Code / Codex plugin that drafts your X and Reddit posts from your PRs and the day's hot threads — drafts only, never posts for you.
More in the works — built in the open, shipped fast.
