Make songs with AI
Connecting a client
How to point Claude, Ollama, LM Studio or any other MCP client at the NØISE song server.
The endpoint is one URL, with no key and no sign-up:
https://noise-app.com/api/mcpIt speaks Streamable HTTP and is stateless: one POST carries one request, and there is no session to keep. That is why it works from anywhere without configuration.
Claude Code
claude mcp add --transport http noise https://noise-app.com/api/mcpClaude Desktop and other config-file clients
Most clients take a JSON block like this:
{
"mcpServers": {
"noise": {
"type": "http",
"url": "https://noise-app.com/api/mcp"
}
}
}Ollama
Warning
Ollama is not an MCP client. It runs models; it does not connect to tool servers. You need a client layer between the two.
Anything that speaks MCP and can talk to a local model will do:
- Open WebUI — point it at your Ollama instance and add the server under its tool settings.
- LM Studio — has MCP support and runs models itself, so it can replace Ollama entirely.
- Cherry Studio — a desktop client that takes both an Ollama endpoint and MCP servers.
mcphost— a command-line host if you would rather stay in a terminal.
Whichever you pick, the server URL is the same.
Which model
Any model that can follow a tool schema will work. What differs is how often it is right about the song.
A larger model needs fewer correction rounds; a 7B model will usually produce a valid song on the first try and a correct one after you fix a chord or two. The server helps either way — it will not let a broken song through — but it cannot invent knowledge the model does not have.
Checking it works
Ask your client to call song_format. You should get back a few pages describing the shape of a song, every legal style, groove, kit and tuning, and one worked example. If that arrives, everything else will work.