JitAPI

The standard approach to giving an LLM access to an API is to dump the entire OpenAPI spec into the context window. For a small API, that works. For something like the Spotify API or Stripe — thousands of endpoints, deeply nested schemas — it blows up. The model either chokes on the context length or hallucinates endpoints that don't exist.

JitAPI takes a different approach. Instead of loading everything upfront, it uses semantic search over endpoint descriptions and a dependency graph that understands which schemas reference which. When an LLM needs to "get the current user's top tracks," JitAPI finds the relevant endpoint, pulls in only the schemas it depends on, and hands back a minimal, accurate spec.

The result is an MCP server you can point at any OpenAPI spec. Register it once, and LLMs can discover and call endpoints on the fly — multi-step workflows included. I use it daily in Claude Code to hit APIs I'd otherwise need to read docs for.

Published on PyPI. Works with Claude Code and Claude Desktop out of the box.