roon-mcp
MCP server to control the Roon music player, enabling AI agents to search music, manage playback, and adjust volume across zones.
README
roon-mcp
An MCP server to control the Roon music
player, so an AI agent can find music and start playback. Built on the official
Roon Extension API (node-roon-api + node-roon-api-transport /
node-roon-api-browse).
"Play a female vocal trance mix on Roon (Zone: This Computer) — 2000s Dutch trance vibes, the kind of set you'd hear at Amnesia in its prime. Start with something like 4 Strings' Take Me Away."
Claude searches across your library and streaming, hand-picks the tracks, and queues them — all from a single natural-language prompt.
Requirements
- Node.js 20+
- A running Roon Core on the local network to pair with
- git on the install host — the
node-roon-api*dependencies are published on GitHub (not npm) and are fetched via git URLs during install
MCP Client Configuration
Add this to your MCP client config. npx fetches the package on first run:
{
"mcpServers": {
"roon": {
"command": "npx",
"args": ["-y", "roon-mcp"],
"env": { "ROON_DEFAULT_ZONE": "Office" }
}
}
}
On first launch, open Roon → Settings → Extensions and enable Roon MCP to pair. Pairing status is logged to stderr; stdout is reserved for the MCP protocol.
Global install (optional)
npm install -g roon-mcp
{
"mcpServers": {
"roon": {
"command": "roon-mcp",
"env": { "ROON_DEFAULT_ZONE": "Office" }
}
}
}
Configuration
| Env var | Purpose |
|---|---|
ROON_DEFAULT_ZONE |
Optional fallback target for play_now / enqueue_and_play when no zoneId is given — a zone/output id or a display-name substring. If unset, the server falls back to the only zone, an Office zone, or the currently-playing zone; if it still can't decide it returns ZONE_AMBIGUOUS so the agent can ask. |
Tools
| Tool | Purpose |
|---|---|
list_zones() |
List playable zones/outputs (id, name, state, output ids). |
search_music({ query, type?, limit?, includeStreaming? }) |
Resolve a text query into ranked browse candidates (opaque, session-scoped item keys). type (artist/album/track/genre/playlist/radio) restricts the category; for non-genre types an empty typed search broadens to all categories. See Streaming search for type:"genre"/type:"artist" and includeStreaming. |
get_tracks_for({ itemKey, limit? }) |
Expand an artist/album/genre/playlist candidate into concrete playable tracks. |
play_now({ zoneId?, itemKey, shuffle? }) |
Immediately play one search candidate; zoneId optional (defaults as above). |
enqueue_and_play({ zoneId?, itemKeys, shuffle? }) |
Build an ad-hoc queue from curated item keys and start it (replaces the zone's queue); reports queued/skipped. |
now_playing({ zoneId? }) |
Snapshot of the zone's current track — state, title, artist, album, seek position. title/artist/album are undefined when nothing is playing. |
control_playback({ zoneId?, action }) |
Run a transport verb: pause / resume / next / previous / stop. |
set_volume({ zoneId?, level }) |
Set the zone's volume to level percent (0–100). Rescales to each output's native range; incremental outputs are reported as skipped. |
mute({ zoneId?, muted }) |
Mute (muted: true) or unmute (muted: false) every output in the zone. |
Streaming search (genre and artist)
Genres don't appear in Roon's flat search hierarchy, so search_music({ type: "genre" })
is handled specially: the server walks Roon's dedicated Genres tree (cached per
session) and returns the nearest-match genre nodes by fuzzy score, with their parent
path in the subtitle — e.g. "Psychedelic Trance" yields Psytrance / Trance. It
never silently broadens to artists/albums. These candidates are library-scoped
(genres present in your collection, including TIDAL albums you've added). Expand one
with get_tracks_for to get a cross-album mix of that genre.
Artist search is library-scoped too: search_music({ type: "artist" }) returns artists
present in your collection. When the top-ranked artist has no library content — e.g. a
node with subtitle: "0 Albums" because all albums were removed, or an artist not in
your library at all but available on TIDAL — the result's message field carries a
hint so the agent can opt in to the streaming path (see below) without having to
discover the dead end at get_tracks_for time.
Pass includeStreaming: true (meaningful for both type:"genre" and
type:"artist") to also pull a track mix from streaming services for discovery beyond
your library:
- For
type:"genre", the server takes the genre-relevant albums the flat search surfaces and samples tracks across them. Library genre nodes come first, then ready-to-play streaming tracks (each atrackcandidate, source groupStreaming) appended after. - For
type:"artist", the server runs a track search and filters to entries by that artist (matching the track's subtitle, which Roon uses for the artist on track rows). Library artist candidates come first, then streaming tracks by the same artist appended after. This is the path that unblocks "play/queue multiple Artist songs" for an artist with no library content: the streaming tracks are queue-playable directly viaenqueue_and_play. Results are best-effort — a small fraction may be features or compilations; the agent's per-artist cap / dedupe handles that.
Default is false (library only).
Cost: with
includeStreamingon, each sampled album/track re-navigates the flat search, so an opt-in streaming search does a handful of extra browse round-trips.
Now playing & transport
now_playing({ zoneId? }) returns a structured snapshot (state, title, artist,
album, seek position) so the agent can confirm what's on and where before
running a transport verb. The server subscribes to Roon's zone-state stream
once per paired Core and waits for the next Changed event after play_now
/enqueue_and_play/control_playback — so nowPlaying in the response
reflects the new track rather than the one that was playing before the
action. The wait times out fast on a slow Core (2 s) and falls back to the
cached snapshot, so a hung subscription can never delay an MCP call.
control_playback takes one verb at a time
(pause, resume, next, previous, stop) — there is no compound
"pause and skip." For "louder" / "softer" without a number, set_volume is
absolute, so the agent should ask for a target percent or apply a default
delta; volume isn't reported in now_playing. Volume and mute fan out to
every output in the resolved zone and rescale per output, so a single
set_volume({ level: 50 }) works correctly across a grouped zone with mixed
dB / numeric devices; incremental outputs (IR blasters and the like) are
reported as skipped in the result.
Assumptions
- Core language: English. Category/action label matching (
Artists,Play Now,Top Tracks, …) assumes an English Core. - Sources: local library + TIDAL. What searches surface reflects this; results differ on a local-only Core.
- Queue: replace.
enqueue_and_playstarts a fresh queue rather than adding to whatever is already playing. - Curation is agent-side. Dedupe / cap-per-artist / ordering / trimming stay in the agent; the server has no curation logic.
There is no save_playlist tool: Roon exposes no official playlist-write service, so
durable playlists are out of scope. Curated playback is delivered by enqueue_and_play
(an ad-hoc, in-the-moment queue).
Logging
Every Roon API call (browse, load, get_zones, change_settings) emits one
structured line to stderr (stdout stays reserved for MCP JSON-RPC):
[roon-call] {"t":"2026-06-19T18:00:00.000Z","lvl":"info","op":"browse","ms":12,"params":{"hierarchy":"search","item_key":"…"},"result":{"action":"list","count":7}}
Failures log lvl:"error" with the mapped error code, and retries surface as repeated
lines for the same op. Grep stderr for [roon-call] to trace a flow.
Development
npm install # pulls node-roon-api packages from RoonLabs' GitHub
npm run build
npm test # builds, then runs node:test
npm run dev # tsc --watch
npm run typecheck
Run the built server directly:
npm start
Integration smoke test
scripts/integration.mjs spawns the built server over stdio (like a real MCP client)
and runs list → search → expand against your Core. Audible steps are opt-in so it never
blasts music by accident:
npm run build
node scripts/integration.mjs # list_zones + search only (read-only)
ROON_PLAY=1 node scripts/integration.mjs # also play_now a shuffled genre
ROON_ENQUEUE=1 node scripts/integration.mjs # also enqueue a few curated tracks
# overrides: ROON_ZONE="Office" ROON_QUERY="Tycho"
Enable Roon MCP under Roon → Settings → Extensions on the first run (the script
retries list_zones for ~45s while it waits to pair).
Releasing
npm run release -- <patch|minor|major>
Requires a clean worktree on main and a matching ## [<version>] entry in
CHANGELOG.md. The script runs the tests, bumps the version, tags the
commit, and verifies the package with npm pack. It then prints the push and
npm publish commands to run.
License
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.