Instead of one MCP server per service, you hand it a specification URL and the endpoints described there become callable — which is the difference between waiting for someone to write a connector and using the API this afternoon.
A LangChain-based server that loads OpenAPI specifications at runtime and turns them into callable tools, reachable either from an MCP client over SSE or from a Langflow canvas through its bundled components. It also runs as a plain CLI, which is the quickest way to see whether a given spec parses before wiring anything up.
- Dynamic loading of an OpenAPI specification from a URL, so an API becomes usable without a purpose-built server for it
- An MCP server mode over SSE — `uv run hippycampus-server --transport sse --port 8000`, reachable at `http://localhost:8000/sse`
- A CLI mode, `uv run hippycampus-cli`, for debugging spec parsing and agent interaction before involving a client
- Two Langflow components: OpenApi Service for loading a spec, and Hippycampus MCP Server for connecting over SSE
Python 3.12.9 and the UV package manager; create a virtual environment, then `uv pip install -e .`, and `uv pip install langflow` if you want the visual side. A Google AI Studio API key in `GOOGLE_API_KEY` — authentication errors almost always mean it is unset. For Langflow, export `LANGFLOW_COMPONENTS_PATH` pointing at the repository's `langflow/components` directory, start the MCP server before Langflow, and restart Langflow if the components do not appear. Specs in the wild are not always correct: the official XKCD swagger declares its comic_id as a number rather than an integer, and the repository ships a corrected copy in its test folder.