Most of the work in wrapping an API as an agent tool is writing down the shapes that someone already recorded in Postman: the parameters, the headers, the response body. Generating from the collection means the tool matches the request that was actually tested, and switching from LangChain to the OpenAI function format is a rerun rather than a rewrite. The generated code is a starting point you review, not something to paste unread.
An MCP server over Postman's Tool Generation API. Point it at a request that already exists in a Postman collection and it returns code for calling that endpoint as an agent tool, shaped for the framework you name.
- Code generated from one Postman request, identified by its collection id and request id — `generate_ai_tool`
- A choice of output language: `javascript` or `typescript`
- A choice of target framework: `openai`, `mistral`, `gemini`, `anthropic`, `langchain` or `autogen`
- Type definitions for the request and response, error handling, the API call itself, function definitions and usage documentation, generated together rather than left as an exercise
A Postman API key in `POSTMAN_API_KEY`, and the collection id and request id of a request in the Public API Network. The npm package name is `postman-tools-server` (0.1.0) and it is marked private, so build from a clone with `npm install` and `npm run build`, then point your client at the built `index.js`.
One command plus a key — npx -y github:giovannicocco/mcp-server-postman-tool-generation, then supply credentials