Stop losing what your agents learn.
AutoSkills watches your Cursor, Claude Code, and Codex sessions and turns what they learn into reviewed, committed skills — AGENTS.md, CLAUDE.md, Cursor rules — written for you. Nothing ever leaves your machine.
the loop
how it works
- capture
adapters read the session stores your tools already write — no hooks to configure, no proxy in the way.
- distill
an llm pass hunts for corrections, repeated rediscovery, and failure-then-fix sequences. no evidence, no suggestion.
- review
one to five suggestions a day, each with the transcript excerpts that justify it. accept, edit, or reject.
- write
accepted skills are compiled to every format your tools read, inside managed markers that never clobber your edits.
- verify
skills earn their place — if the mistake they address keeps recurring, autoskills suggests pruning them.
the review
every suggestion carries its evidence
Use pnpm, never npm — preinstall hook redirects
you corrected the agent twice this week; the repo's preinstall hook rejects npm and the agent keeps rediscovering it the hard way.
evidence — cursor
user: "no — this repo uses pnpm, the npm install you just ran got redirected"
receipts
privacy you can check, not trust
local only
your transcripts, code, and skills never leave your machine — period.
bring your own key
your transcripts go to your llm endpoint — yours, your company's gateway, or local ollama — never through our servers.
open source
every line auditable. the telemetry payload is printed before it is ever sent — and it is off by default.
- AGENTS.md
- CLAUDE.md
- .cursor/rules
- SKILL.md
one canonical skill, compiled to every format your tools read.
waitlist
the local CLI is coming first — open source. the hosted team inbox comes after.
no backend yet — the button opens your mail client