MCP server support

[ranfdev]

LLMs now have the ability to use tools through MCP servers (http servers following some rules).
By implementing a MCP server in Biblioteca, LLMs could then interact with Biblioteca to receive up-to-date documentation about all the GNOME libraries, plus code examples written in the docs.

This could require adding a tool to convert HTML pages to markdown as a dependency, so that we can return the documentation as markdown.

If you are interested, I could try implementing this and open a PR

[sonnyp]

Thanks for the suggestion.

I'm not familiar with MCPs. What exactly would this enable? Do you have a concret example?

I suspect Biblioteca would have to run in the background then?

[ranfdev]

To use MCPs, you need to use an agentic coding tool, that could be VSCode or Claude Code or gemini-cli, etc...

An MCP server exposes a set of tools that can be used by an LLM.

For example, the LLM may output a specific message type, like

{type: "tool", name: "Biblioteca.search_doc", arg: "AdwNavigationView"}

Your agentic coding tool will parse that response, and call the specific MCP server you have configured. It will then pass back the result as

{
  type: "tool",
  response: "doc about NavigationView"
}

Well the syntax is not exactly like that one, but the point is, you can expose a set of functions to an LLM integrated in your editor.

This prevents the LLM from hallucinating too much stuff: it can call the documentation tool to retrieve the relevant up-to-date docs.

This can be supported by Biblioteca by running a HTTP server in the background, or maybe, even just by handling each MCP request from stdin and writing a response to stdout (in theory the MCP protocol supports this).

[sonnyp]

A PoC in Biblioteca and a short demo (to measure impact) would be great.

I prefer http over stdin/stdout for IPC.

[sonnyp]

Ok so MCPs are to "AI agents" what Language Servers are to editors.