Bringing vintage documentation back to life

When people talk about preserving computing heritage, they usually think about hardware, resoldering motherboards, dumping cassettes, saving things from oblivion.

But something essential is often missing: the knowledge.

How do you find, without 30 minutes of searching, the exact syntax of a 6809 assembly routine?
Or that obscure trick for compiling a C program on the MO5?

That’s exactly why I built retrocomputing-ai.cloud: a RAG (Retrieval-Augmented Generation) server designed to help bring these machines back to life — by making them programmable again.

My approach: a very focused development assistant

For now, I’m deliberately focusing on one playground: 👉 the Thomson MO5

BASIC is already well covered everywhere.
But as soon as you venture into:

…information becomes scattered, sometimes contradictory, and often hard to find.

And let’s be honest: I have a MO5 sitting right here… that helps 😄

The idea is simple:

ask a question in plain language,
and let the RAG find the right page, the right excerpt, the right explanation.

Eventually, I’d like to extend the platform to other iconic machines: Commodore 64, Apple II, Amstrad CPC…

What’s in the knowledge base

The documentation covers several layers, kept well separated to avoid mixing everything up:

📄 Official MO5 documentation
Reference PDFs covering the hardware, the system, BASIC, primary sources, digitized and indexed.

⚙️ CMOC
Dedicated documentation for the C compiler targeting the 6809, which I’ve synthesized into a Markdown file to make it more usable by AI tools.

🛠️ MO5 SDK
All SDK .h files are documented in Markdown (prototypes, descriptions, examples). They’re injected into the RAG in a separate category, and installed automatically if you start from the template (make install).

📚 How-tos and patterns
Guides I wrote covering techniques specific to MO5 game development: VBL, RLE compression, collision detection, memory optimizations… The kind of stuff you can’t find anywhere, or that you rediscover the hard way after three lost evenings 😄

A complete toolchain for MO5 development in 2025

The RAG server doesn’t live in isolation. It’s part of an ecosystem I built piece by piece:

🛠️ SDK MO5

A C library optimized for CMOC, with MO5-specific abstractions: sprite management, graphics mode, input handling, etc.
👉 github.com/thlg057/sdk_mo5

📦 MO5 Project Template

A ready-to-use project template with an automated Makefile.
3 commands in a GitHub Codespace, and you’re compiling your first C program for the MO5, with nothing installed locally.
👉 github.com/thlg057/mo5_template

🤖 MCP Server (this site)

The bridge between your coding agent and the MO5 knowledge base. Plug it into Claude Desktop, Cursor or Augment, and your AI suddenly becomes a MO5 specialist.
👉 npmjs.com/@thlg057/mo5-rag-mcp

A word on the technical side (without going too deep into the wiring)

The server runs on a fairly standard RAG pipeline, but designed to stay self-contained and cheap to run.

Currently:

Why this approach?

The model has its limits, it sometimes lacks semantic understanding, and the results aren’t always optimal. But it does the job, and most importantly: no extra costs. Just my VPS hosting, full stop.

Whatever the engine, the goal stays the same: provide a useful and sourced answer, rather than a convincing hallucination.

Use the MO5 index in your IDE in 2 minutes

Thanks to the MCP protocol, your AI can connect to my knowledge base with no complex setup.

The MCP server is published on npm, no need to clone anything. Just add this to your MCP configuration file (Claude Desktop, Cursor, Augment…):

{
  "mcpServers": {
    "mo5-rag": {
      "command": "npx",
      "args": ["-y", "@thlg057/mo5-rag-mcp"],
      "env": {
        "RAG_BASE_URL": "https://retrocomputing-ai.cloud"
      }
    }
  }
}

That’s it. No git clone, no npm install, no path to configure. npx downloads and runs the server on the fly. 🎉

The MCP server is also listed in the official MCP registry (registry.modelcontextprotocol.io), making it discoverable directly from GitHub Copilot and other compatible clients.

What your agent can actually do

Beyond documentation search, the MCP server also exposes build tools that your agent can use directly:

In practice, this means your agent can handle the full cycle: documentation, code, compilation, floppy generation, all without leaving your IDE.

What’s next?

The project moves more by experimentation than by planned announcements.

I keep filling in the documentation as I discover things, sometimes correcting old misunderstandings, sometimes adding clearer examples. On the development side, the SDK now handles sprites and graphics mode properly, I even managed to get a rocket ship firing in vibe coding 🚀

A few pieces are still missing before the Space Invaders is done: joystick support, sound, and the inevitable optimizations that come when someone other than yourself actually tries the thing 😄

But overall, the toolchain holds up. What was still a work in progress a few months ago has become something usable and even fun.

I’ll keep updating the documentation regularly.
All my Markdown files are available alongside the RAG server sources, nothing is hidden, everything can be read, improved, extended.

And if you want to contribute — documents, fixes, ideas get in touch. On my own, I can’t keep this running forever, and I’d love for it to become a genuinely useful tool for retro-computing enthusiasts.

Enjoy 😄

Go further