Giving an AI Real Synth Docs: Building a Surge XT MCP Server

Ask a language model how to recreate a sound and you'll get a confident answer. Ask what one particular knob on Surge XT does and you'll get a confident answer of another kind: assembled from training data, and often wrong in the details that matter.

Surge XT has hundreds of parameters across oscillators, filters, effects and modulation. Nobody keeps that in their head, and no model does either. A smarter model won't fix it. The manual will.

The idea

The Model Context Protocol lets an AI client call tools you define. surgext-mcp turns the official Surge XT docs into something a model can look up instead of remember. It downloads the docs from the Surge repository, splits them along their headings into data/chunks.json, and serves the pieces over MCP.

Four small tools

A common mistake is one giant "ask anything" tool. Small tools give the model something to reason with:

• search_docs(query, top_k?) finds sections by keyword or concept. This is where a model starts when it doesn't know the exact name.
• list_sections() shows the doc hierarchy, so it can see what exists instead of guessing.
• get_section(section) fetches a whole heading path, like Oscillators > Wavetable.
• get_parameter_info(parameter_name) returns the documentation for a single knob.

Search, orient, read, verify. It's the order a person uses a manual, and it turns out to be the order a model does too.

Split on headings

I chunked by the document's own headings, not by character count. A fixed-size split will cut a parameter description in half or glue the end of one section to the start of another. Heading-aligned chunks stay whole, so what a search returns is a coherent answer to a coherent question.

Recreating a patch

I built this to recreate sounds by ear. The repo includes a sample system prompt for it: describe the target sound, have the model look up the relevant oscillator, filter and modulation sections, then propose settings backed by what the docs say. Every suggestion traces to a section instead of a hunch.

It works with Claude Desktop, Cursor and OpenCode. Register the server in each one's MCP config, pointing at dist/index.js. Setup is bun install, bun run build, bun run ingest.

What I took from it

This is retrieval-augmented generation with the ceremony stripped out. No embeddings, no vector store. A well-chunked, authoritative document and plain full-text search were enough, because the corpus is small and structured. If your domain has a manual, wiring it in as tools may be the cheapest accuracy gain you'll find.

The code is on GitHub.