I've been dealing with the MCP side for a while, and I wanted to share what finally came up: mcpify.
OpenAI 相关模型动态已经出现，适合跟踪能力变化、生态影响和后续可用性。
一个名为 mcpify 的新工具简化了使 OpenAPI REST API 可供 AI 代理使用的过程。它允许开发人员通过单个命令集成 API，无需从头开始构建 MCP 服务器。一个显著的特点是其“--lazy”模式，该模式针对大型 API 动态调用必要的工具，从而显著减小工具列表的大小。例如，在 api.weather.gov 的测试中，工具列表从 38,882 个字符减少到 1,741 个字符。该项目已在 GitHub 上发布。
The basic idea is simple: if you have an OpenAPI REST API, it makes it useable by AI agents with a single command without writing a MCP server from the beginning.
But I didn't want to leave it as just turning the endpoints into the tool. I also included things that will come in real use such as Auth, OAuth2, read-only/policy rules, retry, caching, stdio + Streamable HTTP, health check. With mcpify doctor, you can check whether the API is agent-friendly, with mcpify try, you can try tools from the terminal without opening any MCP client.
My favourite part was the --lazy mode. In large APIs, instead of putting the entire tool list on the model, it calls the required tool. In the api.weather.gov example I tested, the tool listing has dropped from 38,882 characters to 1,741 characters.
I also wanted to reduce dependencies as much as possible; the runtime side is based on Python stdlib. There are currently 294 tests and there are true MCP protocol tests on both stdio and HTTP sides.
I published it as an open source. I would especially like to hear it if there is a criticism, a bug or something that you say “you should definitely add this”.
GitHub: https://github.com/furkan708/mcpify