# agent-tuner

> A zero-dependency CLI that stops AI coding agents from repeating mistakes. It
> logs an agent's failed code generations, distills them into project-specific
> rules and regression evals using your own LLM, and auto-updates your agent
> config files (CLAUDE.md, AGENTS.md, .cursorrules, copilot-instructions,
> Windsurf). A self-improving feedback loop tied to your codebase — not a static
> prompt library.

## What it does
- `npx agent-tuner init` — set up in a repo, auto-detect agent config files.
- `log` — record a failed generation (what the agent got wrong + the error output).
- `distill` — group failures and turn them into sharp, durable rules + one regression eval each.
- `apply` — write rules into an idempotent managed block in your agent config.
- `eval run` — re-run the failing prompt with the new rules; exits non-zero on regression (CI gate).

## Who it's for
Small dev teams adopting agentic coding tools (Claude Code, Cursor, Copilot, Codex, Windsurf)
who want their agents to stop making the same project-specific mistakes every sprint.

## Pricing
- Free: 1 repo, $0, all local features.
- Team: $19/mo per repo — sync rules across repos.
- Org: $49/mo — org-wide shared ruleset.

## Key facts
- Zero runtime dependencies; installs via npx in seconds.
- Model-agnostic: works with Anthropic, OpenAI, OpenRouter, Vercel AI Gateway, LiteLLM, Ollama.
- Rules and evals are git-tracked JSON your team commits and owns.
- Differentiator vs prompt-library tools: rules are generated from *your* repo's real
  failures and are continuously regression-tested.

Docs: https://tuner.aiskillhub.info