Paste an OpenAPI document and get MCP tool definitions and a working Python server scaffold. Everything runs in your browser — your spec is never uploaded.
A generated MCP server is a starting point, not a demo. Deploy it, point MCP Showcase at the URL, and get a live playground your customers can actually try.
Nothing is uploaded. The conversion happens in your browser, which matters when the spec describes an internal API.
Path, query and body parameters are flattened into one JSON input schema per operation, with descriptions carried across.
You get the MCP tool definitions as JSON and a FastMCP Python server ready to fill in auth and deploy.
Paste an OpenAPI (or Swagger) JSON document above and this converter maps every operation to an MCP tool: the operation becomes the tool name, its summary becomes the description a model reads, and the path, query and body parameters are flattened into a single JSON input schema. You get the tool definitions on their own, plus a FastMCP Python server built around them that runs as soon as you add authentication.
The whole conversion happens in your browser. No request is made, nothing is uploaded, and nothing is stored — which matters, because an OpenAPI document is very often a description of an internal API that has no business being pasted into somebody else's server. You can verify that in your browser's network tab.
Converting all of an API is the obvious move and it is usually the wrong one. An MCP server's tool definitions are sent to the model on every single turn, so eighty endpoints become eighty descriptions and eighty schemas in the context window of every request — and the model's ability to pick the right one falls sharply as that list grows.
Start with the handful of operations an agent genuinely needs. Add more only when something turns out to be missing. You can measure exactly what a given tool list costs with the MCP token calculator.
$ref are emitted as a generic object
rather than expanded, because resolving them needs the components section. Flagged in the
output rather than dropped quietly.Running it proves the mapping worked. It does not help anyone else understand what the server is for — and a customer evaluating your MCP integration cannot read a Python file. Deploy it, point MCP Showcase at the URL, and you get a live playground with generated documentation per tool, where a prospect can try the thing in a browser instead of taking your word for it.