Feature Proposal: Add MCP (Model Context Protocol) Server for Interactive Docs and Chatbot Integration

[flefevre]

Summary

I would like to propose the integration of a lightweight MCP (Model Context Protocol) server into MkDocs Material. The goal is to allow AI tools (like Langflow, Open WebUI, etc.) to interact directly with the documentation content and structure — enabling use cases such as contextual chatbots, intelligent search, and dynamic assistants.

What is MCP?

The Model Context Protocol (MCP) is a conceptual interface through which LLM-based systems can query structured documentation, retrieve relevant context, and reason more effectively. It defines a standard way to expose:

• Document sections and their hierarchy
• Page metadata (title, tags, etc.)
• Searchable content
• Context windows (breadcrumbs, relations, etc.)

This protocol would make it possible to use MkDocs-powered sites as intelligent knowledge bases, with minimal setup.

Proposed Implementation (Initial Idea)

• Optional dev server plugin (mkdocs serve) that exposes an MCP-compatible endpoint (e.g. /mcp/context)
• JSON responses structured for LLM consumption (e.g. title, content, path, headings)
• Configurable context depth or filtering
• Optional real-time updates for dev workflows

Use Cases

• Integration with Langflow / Open WebUI / Ollama-based chatbots
• Embedding AI copilots into MkDocs websites
• Contextual documentation Q&A during development
• Enhanced semantic search using AI

Why This Matters

With the growing adoption of AI assistants and custom LLM workflows, documentation is no longer just static content — it becomes an interface. A native MCP integration in MkDocs Material would make this project even more future-proof and open up exciting new developer experiences.

[squidfunk]

Thanks for suggesting! We're currently knee deep in foundational work, but once that's settled, we might consider adding this functionality. It should probably live as a plugin anyway, so maybe somebody else is faster than us ☺️

[ubaumann]

Over the long weekend, I hacked a PoC together. It is far from perfect, but it would be interesting to get some feedback. As this is my first mkdocs "plugin", I am sure there is a lot I can improve.
https://github.com/ubaumann/mkdocs-mcp

[walmsles]

I came here to create a discussion item and found this. I tackled this from a different direction. I have created a project for a stand-alone local MCP server which is capable of loading the live search_index.json file into a local lunr.js instance to enable AI agents to do a search which provides the same output and experience as the live docs site. I have also provided a tool which enables the fetching of the page from the web. Its an early stage working Prototype that plugs in nicely into Calude Desktop or other local MCP setup.

I did this for the Powertools for AWS Lambda documentation to check out how it can empower AI assistants (see here). It also then scrapes the live doc pages and converts to markdown for AI consumption. After building this I thought - this could be useful for mkdocs more generally and I get that its not your core purpose here, so I came here to say - would you be interested if I made this more generic and package it up as an add-on for mkdocs (and found this feature proposal).

Also interested in input from @squidfunk so we can we do this responsibly without annoying the whole world.

Am curious how you see this as a plugin and how that would work - What I have built directly hooks into a deployed site and simply works - I feel there should be some mkdocs integration to enable conformity and approval of using an MCP server to access doc sites (see: not annoying the whole world).

I also wanted to ask if I can integrate your front-end code for re-constituting the index rather than the cobbled version I have now (I was moving fast the other day).

Let me know your thoughts. Happy to collaborate on this idea.

[squidfunk]

As said, the whole team is currently busy with foundational work, but once that's finished we can consider how we can improve LLM support. I'm not sure whether MCP make sense, as we're building a static site, but maybe there is some direction that fits.

[squidfunk]

I also wanted to ask if I can integrate your front-end code for re-constituting the index rather than the cobbled version I have now (I was moving fast the other day).

Take anything you want. It's all licensed under the terms of the MIT.

[walmsles]

I love where you are going with the foundational work! Thanks for the response.