# Introducing MCP Documentation Servers: Real-time Mojo Docs for Your AI Models
Hey Mojo community! Longtime lurker, first-time poster.
I’ve been working on something that I think solves a real pain point, especially for those of us using AI models (like Claude in VS Code) alongside a rapidly evolving language like Mojo.
## The Problem
Mojo is changing fast. New features, updated syntax, better APIs—it’s exciting, but it creates a challenge: **how do models stay current?** By the time documentation makes it into training data, it’s already outdated. And copying snippets from docs into prompts? Tedious and error-prone.
## The Solution
I built a framework for creating **searchable MCP servers** that expose documentation with **hybrid search** (vector + keyword matching). The practical result: your AI models get instant access to up-to-date, relevant documentation through [Model Context Protocol](https://modelcontextprotocol.io/).
Think of it as giving Claude or your model a live, searchable knowledge base that’s always in sync with the actual language and doesn’t overwhelm the context window.
## How It Works
1. **Process documentation** (MDX/Markdown) → extract and chunk it intelligently
2. **Generate embeddings** using MAX’s `sentence-transformers` model
3. **Index with DuckDB** (HNSW for vectors, BM25 for keywords)
4. **Expose via MCP** so models can search and retrieve context
The server is self-contained—once built, it runs anywhere with just Pixi or Python. Perfect for distributing across different documentation sources.
I use the mojo manual, directly from the Modular repo, as the source documentation. I plan to update the database, any time there is a change in the docs.
## Why This Matters for Mojo
Mojo is new and evolving rapidly. By the time you ask a model “How does ownership work in Mojo?” you want the *current* answer, not something from 2024. This approach ensures models always have access to the latest manual.
Plus, hybrid search means better results—semantic understanding *and* exact keyword matching.
## Visit the Repo
Head over to **[github.com/jpotter80/mcp](https://github.com/jpotter80/mcp)** for the full framework and a working Mojo Manual MCP server.
The framework is there for creating new servers, from other docs. However, the mojo-manual-mcp is ready to use right now, no build pipeline necessary.
Setup is straightforward (Pixi recommended):
```bash
git clone https://github.com/jpotter80/mcp
cd path/to/mcp/servers/mojo-manual-mcp
pixi install
```
Then, copy the json config into the VS Code mcp.json file, and hit start server button. You may have to restart VS Code for the config to take effect.
{
“servers”: {
“mojo-manual”: {
“type”: “stdio”,
“command”: “pixi”,
“args”: [“run”, “serve”],
“cwd”: “/absolute/path/to/mojo-manual-mcp”
}
}
}
Since I use VS Code, I have not tested this on Cursor, Claude Desktop, etc. I understand that the json config is slightly different, depending on client. I will be testing those clients, in the future.
James Potter