Summary
The spec stays out of the context window until the model asks.
A large OpenAPI file pasted into a conversation eats the budget and still gets misread. Here it is loaded once under an ID and the model pulls one endpoint at a time, asking for prose rather than schema when that is easier to reason over. Several APIs can be loaded side by side, which is how you get a model to work across two services.
What it is
A Node server that holds one or more OpenAPI or Swagger specifications in memory under IDs you choose, and answers questions about them one endpoint at a time.
What you get
- load_api to bring a spec in from a URL or a file path under an apiId of your choosing
- list_apis for everything currently loaded
- search_endpoints to find endpoints matching a pattern in the path or the description
- get_endpoint_details for one method and path, with a natural flag that returns a human-readable summary instead of raw schema
- The same material as resources: an overview per API, its endpoint list, and a single endpoint by method and path
Requirements
Node, and a clone of the repository built with npm; the built index.js is referenced by absolute path in the client config. Specifications are loaded at runtime by load_api, so nothing is fixed in advance.