[WIP] Create model context protocol for MP's API

[esoteric-ephemera]

Model context protocols (MCPs) are a generally architecture agnostic standard for allowing LLMs to access the key functionalities of an API or other walled-off data (or abstract functionality)

Want to note that a previous effort within LBL was undertaken to write an MCP for MP. This expands on that to encompass all of the API client's functionality.

This still uses the FastMCP package to generate an MCP. There is a 10-line, easy solution for generating an MCP programmatically from an OpenAPI spec. The FastMCP developer recommends this approach only for bootstrapping. Basically LLMs struggle to understand what is important in the API if there is no structure beyond the OpenAPI spec.

A more fleshed-out solution using the (mostly) auto-generated mp_api.mcp.tools is included as well. These basically structure the core functionalities of the API client by both using MPRester's convenience functions, and the search features.

Still needs work and testing

[tschaume]

Awesome! Only have a couple of comments.

Should regenerating the tools be integrated with Github actions and run during a release?

[esoteric-ephemera]

@tschaume Re:

Should regenerating the tools be integrated with Github actions and run during a release?

It's mostly automated but requires some cleanup work. A key thing is ensuring that the typing is correct (FastMCP needs explicit type annotations) and that any new types are imported. Could be done if that's of interest

Before this gets merged, I'd like to test a bit or get feedback from those who are more in the LLM space

[computron]

@esoteric-ephemera how about Thrust 3 of SciDAC (Knowledge) get involved with the MCP testing? Let me know if you want me to mobilize them, happy to do so as a first concrete action item for that team

[esoteric-ephemera]

Thanks @computron! That might be helpful.

Right now, I have it integrated with Claude desktop, but am running into limits testing it for "enterprise" LLMs (e.g., ChatGPT needs a web-accessible MCP server), especially with the limited tokens you get at the freemium tier

Would be good to know if anyone on SciDAC has worked with Llama 3 or another good open source LLM, or even has integrated an MCP with LBL's CBORG