shivamchavan01/mcp_foundry

MCP Foundry takes an OpenAPI specification, the kind of document that describes what an API can do, and turns it into a working MCP server. MCP stands for Model Context Protocol, a way for AI assistants like coding agents to call tools. Instead of just generating code and stopping there, this project also tests every generated tool against a temporary mock version of the API, scores how well each tool performs, and automatically tries to repair the ones that fail, up to three rounds of fixing. The project describes its approach as generate, evaluate with real agent prompts, repair, and evaluate again. Each generated tool gets a visible score rather than a simple pass or fail hidden behind a checkmark, and the project intentionally reports a needs_review status when it cannot confirm a tool works reliably, rather than pretending everything passed. Under the hood, the pipeline parses the OpenAPI spec, spins up an isolated mock server using a tool called Prism so nothing touches the real API during testing, rewrites vague tool descriptions so an AI can tell similar tools apart, generates the actual server code using the FastMCP framework, runs the agent style evaluation, and if successful, registers the finished server into a coding tool called Codex. To use it, you need Python 3.11 or newer, Node.js and npm, and an API key from one of the supported AI providers, which include Mistral, Groq, Nvidia, and Google. There is a web interface where you click a button to start generating a server, and it also exposes a plain API you can call directly with curl commands. Generated servers, along with their evaluation results, are saved to files on disk. The README notes clear limitations: registration only works on the same machine where the server was generated, job status is stored in memory so restarting the backend loses progress, and results depend on the chosen AI provider staying available. The README states this is a hackathon project and asks users to add a proper license before publishing or distributing it, so its licensing terms are currently undefined.