what is pantry fr?

pantry is a storage shelf for reusable code snippets that AI agents can fetch on demand. Instead of an AI assistant re-deriving the same logic every time it runs, you save that code once as a "recipe," and any tool can retrieve the exact saved version whenever needed. The key promise is that pantry stores and returns your code but never runs it, the calling tool decides whether and how to execute it. A recipe is a named JavaScript function with metadata: a description, an input schema, capability tags, and a version. You push recipes to the store via a simple API or a command-line tool. When an agent needs a piece of logic, it lists available recipes, fetches the one it wants, reviews the code, and then runs it in whatever environment it chooses. The store is private by default, scoped to an owner via a bearer token, though authors can optionally mark recipes as shared so other owners can discover and read them. This is designed for teams running AI coding agents or orchestrators that benefit from a shared library of vetted, reusable functions. For example, if you have a text- transformation utility your agents use repeatedly, you save it once as a recipe. Any connected tool, an MCP client like Claude Desktop, the Pi agent framework, or a plain curl command, can fetch that same recipe. The capability tags help agents reason about what a recipe does before fetching it. The project runs on Cloudflare Workers and uses D1 for storage, which keeps it lightweight and globally distributed. Security is straightforward: every request requires a bearer token, owner scoping prevents cross-access, and the server fails closed if misconfigured. Notably, the demo runner that executes fetched code is explicitly called out as not a sandbox, it does some basic scanning but the project is honest that running untrusted code is the caller's responsibility.