Written, run,
pinned, re-run.
A recipe is a small, working integration of two or more libraries, plus the tests that prove it works.
A recipe is written for one job
For example: Supabase auth in Next.js middleware. One task, the files it needs, and tests that exercise it.
It runs in a clean sandbox
Fresh install, exact versions, the full test suite. If anything fails, the recipe is not published.
Every dependency is pinned
The recipe records the exact version of each library it passed on, so your agent knows what it is copying.
New releases trigger a re-run
When a pinned library ships, the recipe is tested again. Passing runs update the receipt. Failures mark it red and queue a repair.
Your agent pulls it over MCP
Claude Code, Cursor and other MCP clients search by stack and task, then get the code with its versions and receipt.
Written by a test run, never by hand
Status (passed, flaky or failed), the versions tested, how many tests ran, the runtime, and when.
If it is not listed, it is not tested
We start with common TypeScript and Python web stacks. A passing receipt covers the listed versions at the listed time, nothing more.
Hear when Pro opens.
The free server is live now. Join for Pro plan and new-stack announcements. No newsletter.