A Looperman MCP server for Claude and Cursor
LoopHaven ships an optional MCP server with six tools: search loops, search acapellas, read one acapella, get a preview URL, export a ZIP, and toggle a favourite. The agent receives JSON, never HTML — so it cannot invent a filter value, and it cannot follow one of the sixteen download links on a detail page and grab the wrong loop.
Configuration
The MCP server is a thin client over the local HTTP API, so the API has to be running first. Point the agent at the same port.
# terminal 1
python -m uvicorn loophaven.api.app:app --host 127.0.0.1 --port 8000
{
"mcpServers": {
"loophaven": {
"command": "loophaven-mcp",
"env": { "LOOPHAVEN_API": "http://127.0.0.1:8000" }
}
}
}
That block goes in the MCP configuration of the client — Claude Desktop, Cursor, or anything else that speaks the protocol. No API key, no account, no remote service.
The six tools
| Tool | What the agent gets |
|---|---|
search_loops | Results with title, BPM, key, genre, author, duration and MP3 preview URL |
search_acapellas | The same, plus the licence wording the page displays |
get_acapella_detail | One acapella, with its usage terms in full |
get_preview_url | A playable MP3 URL for one item |
export_loop_zip | A 2–4 loop pack, with a README recalling the licence |
toggle_favorite | Favourite state on the account, when a session is open |
Why JSON matters more than it sounds
Handing a language model the raw listing page is not a shortcut. Every card
on the page shares the same parent element, and a detail page carries sixteen
/getfiles/ links — fifteen of which belong to recommended loops.
A model reading that HTML will confidently attribute the wrong BPM to the wrong
loop, and nothing in the output will look wrong.
Resolving the vocabulary before the model sees it removes the whole class of
error: "{'genre': {'id': 54, 'label': 'Trap'}}" is either right or
absent, whereas gid=54 in a URL is unverifiable to a reader.
The etiquette the tools assume
A tool that can download is a tool that can download ten thousand files. The repository includes a skill file that tells an agent how to behave, and the constraints are deliberately blunt:
- Preview before download. Auditioning costs nothing; downloading costs quota.
- Read the licence label. Some acapellas are non-commercial only. The tool returns the wording; the agent should quote it rather than summarise it.
- Never bulk-harvest. A search returns a page, not a catalogue.
- A
429means wait. Not retry, not switch account.
What it is not
It is not a plugin for your DAW, and it does not place audio on a timeline. The agent finds and fetches; arrangement happens where you work. See REAPER and other DAWs for the half-step that gets a file into a project folder.
Next: the filter reference, so you can tell the agent what to ask for.