LMMS MCP
Provides tools for creating, inspecting, revising, routing, automating, and rendering editable LMMS music projects and local WAV loops using an installed LMMS renderer.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
LMMS MCP

Write game music with an AI agent and get engine-ready files back. The agent composes with chords, drum grids and melody lines, chooses from the whole stock LMMS sound library, mixes with real effects, checks its own work by measurement, and exports seamless loops, intros, adaptive layers and stingers with a manifest. The MIT-licensed server works with saved .mmp and .mmpz files and calls the installed LMMS renderer without opening the editor.
Windows setup
Install Node.js 20 or newer. From a checkout run ./setup-windows.ps1. It installs the Node dependencies and uses WinGet to install stock LMMS 1.2.2 if needed. For manual setup, run npm ci and install LMMS. The package contains our server and demo, and setup obtains LMMS from its official distribution.
Configure a stdio MCP client with command node and the absolute path to server.mjs as its argument. Set LMMS_WORKSPACE to an existing directory for trusted sessions and renders. The package requires Node.js 20 or newer. It is tested with LMMS 1.2.2 on Windows x64; CI uses Node.js 22, and the server and unit tests also pass on Node.js 20.
Variable | Purpose |
| Folder for projects, renders and exports |
| Path to |
| Custom preset, sample and plugin folders |
| Path to ffmpeg for OGG output (defaults to |
| Fixed render limit; the default scales with song length |
| Renders run side by side for stems and layers (1 to 4, default 3) |
| Folder of open SFZ libraries for |
To run the published package, use npx -y @sarutobi-sasuke/lmms-mcp. For an install that follows a fixed GitHub release, run npx -y https://github.com/SarutobiSasuke8/lmms-mcp/archive/refs/tags/v0.9.1.tar.gz. Configure your client to launch that command over stdio with LMMS_WORKSPACE set. Install LMMS separately.
On connection, the server sends a short workflow brief through MCP initialization. Call get_started for the same guidance, workflow entry points and local configuration. The MCP tool list documents exact parameters, including optional live tools when configured.
Stem, layer and full-song renders can take a minute or more. Give your MCP client a tool timeout of at least five minutes.
Related MCP server: kuruitsugizaki
Tools
The saved-project tools work with stock LMMS. Claude's game-audio suite added chord, melody and drum-grid inputs, measured sound search, batch edits, mixer processing, audio analysis and adaptive game exports. Optional components extend the same project workflow with Fathom synthesis, VST3 stems and live controls.
Area | Tools |
Compose |
|
Sounds |
|
Mix |
|
Listen |
|
Deliver |
|
Read |
|
Start fast |
|
Write tunes |
|
Whole pieces |
|
Solo lines |
|
DJ |
|
Scores and performance |
|
Phrase library |
|
Browse sounds |
|
See a song |
|
Check |
|
Balance |
|
Loops |
|
Audition |
|
External instruments |
|
A beat has 48 ticks and a 4/4 bar has 192. Bars start at zero. Files use simple names in the workspace, and existing files are protected from overwrites unless a tool says otherwise.
Composing
Every pattern takes exactly one of four inputs.
{"name": "Beat", "bar": 0, "bars": 4, "steps": {"grid": "X---x---X-x-x---"}}
{"name": "Tune", "bar": 0, "bars": 4, "melody": {"line": "D5:3 F5:3 A5:6 G5:2 F5:2 | G5:3 F5:3 D5:6 r:4"}}
{"name": "Chords", "bar": 0, "bars": 8, "chords": {"progression": "Dm Bb F C", "style": "arp", "order": "updown", "rate": 2}}
{"name": "Raw", "bar": 0, "bars": 1, "notes": [{"tick": 0, "length": 44, "midi": 62, "velocity": 90}]}Step grids tile to fill the pattern.
Xis an accent,xa hit,osoft,ga ghost note and-a rest. The default is 16 steps per bar.Melody lines are
NOTE:stepstokens, with an optional:velocity.r:4is a rest and[D4,F4,A4]:8is a chord. The line repeats to fill the pattern.Chord progressions take one chord per bar (
Dm:2lasts two bars) and loop.styleisblock,pulse,arporbass. Voice leading keeps chords in one register.resolve_chordsshows the notes before you commit.swing,humanize,stepsPerBarandtransposeare available on the generated inputs.
LMMS sizes a clip from its last note, so empty bars at the end of a song normally vanish. create_project pins the song to lengthBars (by default the last declared pattern bar), and set_song_length changes it later. A 45-bar loop stays 45 bars.
copy_bars repeats a section across tracks. add_track and add_pattern extend a saved project.
Use edit_sample_clip for a saved project's native audio lanes. Select a sample track by name and a clip by its start tick. place creates the track if needed and takes a workspace WAV, Ogg or FLAC filename and a timeline length. move and duplicate take a destination tick, relink takes a workspace audio filename, and resize changes the saved clip length in ticks. Keep the referenced audio files with the project and relink if their absolute paths change.
Sounds
list_samplesbrowses the stock sample library (about 1,000 files). Use a result asfactorySampleon avoice: "sample"track, for exampledrums/snare01.ogg. Workspace WAVs work too, including one folder down, so a kit made byexport_one_shotscan be used directly.voice: "preset"loads a factory preset at creation.apply_presetchanges it later. Both read current and older.xpffiles and ZynAddSubFX.xizpatches (954 in stock LMMS).Some presets are saved at another base note and sound up to two octaves away from the written pitch. The default
tuning: "concert"corrects the track so written notes sound at pitch. The result reports the measured offset and level. Passtuning: "preset"to keep the preset as saved.search_soundstakes arole(bass, lead, pad, pluck, bell, keys, strings, brass, organ, drum, fx) and free words. Measured entries carry tags such asbright,percussiveorslow-attack, with level and tuning. The package ships measurements for stock LMMS 1.2.2. Runbuild_sound_catalogueto measure a different installation.preview_soundrenders a short audition of a preset, stock sample or Fathom patch and reports its level and detected pitch.
Mixing and effects
edit_effect adds effects to an instrument track or to a mixer channel such as Master. list_effects returns every parameter with its range.
Plugin | Parameters |
|
|
|
|
|
|
|
|
| Gain and pan, bass boost, stereo width |
| Added at flat defaults; shape them in LMMS |
Delay times are worked out from the project tempo when the effect is written. Update the effect after a tempo change. Echo and reverb tails keep ringing after a track goes silent.
apply_edits runs many edits on one project and saves once. Use it for a whole mixing pass. If any edit fails, nothing is written. Pass overwrite: true to update a project in place.
{
"project": "theme.mmp", "output": "theme-mix.mmp",
"edits": [
{"op": "set_track_mix", "args": {"trackName": "Bass", "volume": 112.5}},
{"op": "edit_effect", "args": {"trackName": "Lead", "operation": "add", "plugin": "delay", "wet": 0.25, "params": {"time": "1/8d", "feedback": 0.4}}},
{"op": "edit_effect", "args": {"bus": "Master", "operation": "add", "plugin": "limiter"}},
{"op": "set_master", "args": {"volume": 70}}
]
}Volumes accept fractions. automate_track covers volume, pan, pitch, filter cutoff and resonance, with step, linear or smooth curves. solo is honoured by renders.
Listening by measurement
An agent cannot hear. analyze_audio reports what it needs to judge a mix: integrated loudness in LUFS (ITU-R BS.1770), true peak, RMS, crest factor, clipped samples, DC offset, stereo correlation, energy in six frequency bands, spectral centroid and the loop seam. Give it sections in bars to compare parts of a song, pitch: true to detect notes, and spectrogram to get a PNG returned as an image. It ends with plain-language warnings, for example a section that will thin out in mono.
render_mix reports loudness and true peak with every render. Options:
tail: "wrap"folds the reverb and delay tail back onto the start, so a loop does not cut its own echoes.tail: "keep"lets the tail ring out.targetLufsandceilingDbmaster the file with a look-ahead limiter. Game music usually sits near -16 LUFS.An
.oggoutput name writes OGG Vorbis when ffmpeg is installed.
render_stems renders one aligned WAV per track, in parallel, for songs of any length.
Game export
export_game_audio writes everything a game engine needs in one call.
{
"project": "theme-mix.mmp", "outputDirectory": "theme", "name": "forest-theme",
"introBars": 4, "targetLufs": -16, "ogg": true,
"layers": [
{"name": "rhythm", "tracks": ["Kick", "Snare", "Bass"]},
{"name": "harmony", "tracks": ["Pad", "Arp"]},
{"name": "melody", "tracks": ["Lead"]}
],
"stingers": [{"name": "victory", "startBar": 43, "bars": 1}]
}Loop: the song after the intro, with its tail folded into the start.
Intro: the opening bars as a one-shot that leads into the loop.
Layers: aligned loops per group of tracks for adaptive music. They share one gain and are not limited, so they sum to the mix. The manifest records how closely (
layersSumResidualDb). Effects on Master or shared buses react to the combined signal, so keep effects on individual tracks for layered music.Stingers: short cuts from a bar range with a fade.
Manifest: tempo, bars, loop points in frames, loudness, true peak and a SHA-256 for every file.
Use export_one_shots to process up to 32 mono or stereo 48 kHz, 16-bit PCM WAVs in one call. Give each source a unique output name and choose a new output directory. The tool trims quiet edges, adds short fades, matches sample peaks and writes mono WAVs and a JSON manifest. The manifest reports duration, trim offset and measured levels. Peak matching leaves each take's dynamics intact.
{
"outputDirectory": "rts-pack",
"sounds": [
{"input": "ack-source.wav", "name": "unit-acknowledged"},
{"input": "shutdown-source.wav", "name": "robot-shutdown"}
],
"thresholdDb": -48,
"peakDb": -3
}A sampler track that plays a workspace WAV stores its absolute path. Use LMMS to relink the source if you move the project to a different machine. Stock samples are stored by their library path and move with any LMMS installation.
Fathom Atelier
Set LMMS_FATHOM_WORKLET to the absolute path of a trusted Fathom Atelier checkout's public/worklets/synth-processor.js before launching the MCP. The LMMS MCP package ships the adapter. The synth engine is supplied locally. Use list_instruments to check configuration and discover the stock plugin and preset inventory. The inventory checks files on disk. Third-party plugins loaded inside hosts such as Carla or VeSTige need their own inventory.
Save a full Fathom patch object as JSON in the workspace. Supported model IDs are fathom, vesper and halo. Call render_fathom_track with a saved 4/4 project at a static tempo.
{
"project": "ninja-bomber-source.mmp",
"output": "ninja-bomber-fathom.mmp",
"audio": "fathom-arp-stem.wav",
"trackName": "Fathom Spark Arp",
"patchFile": "spark-dark.json",
"bars": 8,
"cutoff": [
{"tick": 0, "hz": 220},
{"tick": 576, "hz": 375},
{"tick": 1152, "hz": 1300},
{"tick": 1470, "hz": 4850}
]
}The bridge uses the written pitches, note lengths and velocities. It interpolates cutoff by frequency ratio and updates the worklet every 128 samples. The project copy keeps the muted source piano-roll track for editing and MIDI export and plays the rendered audio from a sample track. Rerun after editing notes, tempo, patch or cutoff, with fresh output filenames. The current bridge renders the dry worklet. Add delay and reverb with edit_effect. The audio clip references an absolute WAV path, so relink it when moving the project.
search_sounds lists Fathom patches only when the bridge is configured, or when you ask with engine: "fathom".
Optional Atelier groovebox tools
Link a local checkout of the private Atelier DSP engine with npm link atelier-dsp after registering the checkout with npm link, or set LMMS_ATELIER_DSP to the absolute path to its repository root. Use a branch that contains src/groovebox/index.js. The LMMS MCP server starts without the engine, and the groovebox tools report the local setup needed when called without it.
groovebox_patternbuilds a 16-step project with per-hit velocity and a synth bass lane with accent and slide.load_pad_kitvalidates a kit JSON file inLMMS_WORKSPACEand returns its pad data.render_grooveboxwrites a new stereo WAV or MIDI file inLMMS_WORKSPACE. For sampler pads, pass a map from each pad's asset ID to the workspace WAV file.
External VST3 instruments
Companion tool coverage
Tool or workflow | Status |
FluidSynth and FluidR3 GM | Tested offline from the same editable MIDI as an LMMS cue. Run FluidSynth separately. |
Decent Sampler and Surge XT | Clean Windows setup and combined piano/pad stem render verified. Sample libraries are separate downloads. |
Dedicated electric piano, vibraphone and orchestral libraries | FluidR3 provided a baseline. Dedicated libraries, articulation changes and large orchestral sessions remain to be tested. |
REAPER, ACE-Step, Stable Audio and Lyria | Evaluation candidates outside the verified MCP workflow. |
The MCP package includes the VST3 adapter. Install the optional host dependencies and instruments locally to enable it. Use list_external_tools to see the upstream links, configured plugins and whether the selected Python has the required packages. The tool reports file presence. A render verifies whether a plugin and preset load successfully.
On Windows x64 with Python 3.12 installed, run ./setup-vst3-windows.ps1 from the package or checkout. It downloads pinned portable Decent Sampler and Surge XT plug-ins from their publishers, checks their SHA-256 hashes and creates an isolated host environment. Copy the generated mcp-env.json entries into your MCP client's environment settings and restart the client. The default destination is %LOCALAPPDATA%/lmms-mcp/vst3-v1. Use -InstallDirectory with a new directory for a separate install or to retry a failed setup. The command leaves system plug-in folders alone and sample libraries are obtained separately.
For manual setup, create an isolated Python environment and set its executable and both plugin locations in your MCP client environment:
py -3.12 -m venv .venv-vst3
.\.venv-vst3\Scripts\python.exe -m pip install pedalboard==0.9.26 mido==1.3.3
$env:LMMS_VST3_PYTHON = (Resolve-Path .\.venv-vst3\Scripts\python.exe).Path
$env:LMMS_DECENT_SAMPLER_VST3 = 'C:\path\to\DecentSampler.vst3\Contents\x86_64-win\DecentSampler.vst3'
$env:LMMS_SURGE_XT_VST3 = 'C:\path\to\Surge XT.vst3\Contents\x86_64-win\Surge XT.vst3'Download Decent Sampler and Surge XT from their publishers. The Windows test used the inner 64-bit VST3 binaries. Open the desired Decent Sampler library in a VST3 host and save its state as a .vstpreset. Copy that preset to the workspace and keep the library at the location the preset expects.
To skip the host step, use open SFZ libraries: setup-libraries-windows.ps1 fetches a piano, upright bass, saxophone and jazz drum kit, convert_sfz turns any of them into a .dspreset, and render_vst3_stems plays it through decent-sampler with the dspreset option. See docs/sampler.md.
Export an editable LMMS arrangement with export_midi, then pass the MIDI filename, render duration, and named track assignments to render_vst3_stems. Assign each stem a configured instrument, an optional workspace .vstpreset, normalized parameter overrides and gain in dB. The adapter returns aligned 48 kHz stereo WAVs, a stereo mix and a manifest with the assignments and frame count. Choose a duration that covers the notes and any desired release tail. The output is a fixed-length stem render. For a game loop, set the intended loop length and check the seam before shipping.
Use list_effects and edit_effect for LMMS post-production. You can place bounced WAVs with edit_sample_clip and apply a master-chain compressor, limiter or reverb. For individual mixer buses, load each bounce as an AudioFileProcessor instrument track and route it with route_mixer_bus. Use analyze_audio to check levels and the seam after processing. The external VST3 adapter currently hosts instruments. The LMMS effect tools handle the post-production chain.
Composer profiles
Search for victory, whole-tone, counterpoint, jrpg, minimalism, folklore, sparse hub or orchestration with list_composer_profiles. The 11 profiles include Grieg’s modal drone, Satie’s sparse accompaniment and Rimsky-Korsakov’s register exchanges. Pass a result's ID to create_composer_sketch with an output filename, MIDI root, tempo and bar count. The sketch works with stock LMMS and returns a suggested Fathom patch ID for its motif track. Open it in LMMS or render it with render_mix.
The profiles are original studies in musical techniques. jrpg-battle references Nobuo Uematsu and additive-pulse references Philip Glass’s additive process. The sketches contain original notes and stock LMMS sounds.
Game sound effects
Use list_samples for installed source sounds and preview_sound to audition them. Layer synthesis and recordings in editable tracks using create_project and edit_sample_clip. Shape the result with edit_effect, render variations, then run export_one_shots to trim, fade and peak-match PCM WAVs and produce a batch manifest. Keep source URLs, licences and attribution with third-party assets. External library ingestion, licence-aware search, automated attribution and direct game-engine import are roadmap work.
Codex and Claude Code skill
Copy the LMMS game audio skill with its directory into .agents/skills/ for Codex or .claude/skills/ for Claude Code in your project. Configure the MCP separately using the setup above.
Limits and roadmap
The project builder creates 4/4 sessions at a static tempo. Tempo and master automation are edited in LMMS.
OGG output needs ffmpeg. Without it the tools write WAV.
Effect parameters written by
edit_effectare kept in the project. If the project is re-saved in the LMMS editor,inspect_effectsstill lists the effects but no longer shows those parameters.SoundFont (
.sf2) and VST instruments are not created by this server yet. Stock LMMS for Windows ships no SoundFont.The pitch detector that drives concert tuning reads one clear fundamental. Check
preview_soundfor presets with strong sub-oscillators or detuning.Next: SoundFont tracks, EQ band controls, tempo and master automation, sidechain ducking, and a macOS and Linux test pass.
Keep the native project as the editable master. MIDI exports instrument patterns and non-overlapping Beat/Bassline clips at a static tempo. Drum steps export as sixteenth-note triggers. Synth patches, effects, audio clips and automation stay in the native project. The default pitchMode: "written" preserves the piano-roll notes. Select "playback" to apply the instrument base note and master pitch offsets. Plugin transposition and pitch modulation remain in the native project.
Each stem is rendered with the other tracks muted and fitted to the mix length. Effects that depend on several tracks playing together can sound different in solo renders.
Live session control
The optional Windows transport adapter uses a separately built LMMS Agent Bridge. Launch LMMS with its bridge, then set LMMS_LIVE_PIPE=lmms-agent-bridge when starting this server. Where enabled, live_transport accepts play, pause, stop and seek. The tool returns the bridge's playback state. The upstream playback-volume cap can lower the master fader and mark your open project as modified when you press play. Saved-project edits and exports work with stock LMMS.
With the same bridge enabled, use live_inspect to read project state, arrangement and notes. Pass the returned project token to live_update_track, live_update_clip, live_replace_notes or live_update_effect. Take track and clip indices from a fresh arrangement read, particularly after adding or removing tracks in LMMS. Changes appear in the open project and can be saved in LMMS.
Use live_parameters to discover current control names, values, ranges and session-local model IDs. Set target_type to song, track, instrument, mixer or effect. Supply target_index for every target except song. For an effect, also supply effect_index and effect_target_type (track or mixer). Get the current token with live_inspect and refresh discovery after changing the project or replacing a plugin.
Pass the current project_token and a list of { model_id, value } pairs to live_set_parameters. Use the discovered native units and ranges. The bridge validates the entire batch before applying it and returns the resulting parameter values. Changes stay in the open session until saved. This works for controls exposed by the Windows LMMS Agent Bridge. External Surge XT and Decent Sampler instruments use the offline VST3 renderer.
Use live_create_track for instrument, sample, automation or Beat/Bassline tracks. Create a linked curve with live_add_automation, inspect it with live_read_automation and replace or upsert points with live_edit_automation. The helper links to track volume, panning or pitch.
For your own synth patches, place a compatible LMMS .xpf in the workspace and call live_load_instrument_file on an instrument track. For voice clips, guitar chops and game sound effects, create an AudioFileProcessor track and load a WAV, Ogg or FLAC file from the workspace. Use a note pattern to sequence the sample. SoundFonts and other native preset formats require a compatible destination instrument.
Verification
npm testruns 13 unit tests for the analysis, composing, effect and catalogue code. It needs no LMMS.node verify-game-audio.mjsdrives the server over stdio with LMMS 1.2.2 installed. It checks composing input, pinned song length, preset tuning, older presets and ZynAddSubFX patches, batch edits, effect tails, solo, stems, loudness mastering, analysis with a spectrogram, arrangement edits, the game export and sound discovery.node verify-exports.mjschecks clip edits, MIDI readback, effect bypass and stem duration. Pass a native MMP path to exercise a different project with instrument patterns at the start of the arrangement.npm run verifyruns both LMMS scripts.
Set LMMS_FATHOM_WORKLET and LMMS_FATHOM_TEST_PATCH to the local worklet and a full patch JSON to also verify Fathom cutoff rendering and rebouncing.
See the comparison notes for tested workflows, a survey of other LMMS servers and known gaps.
Run npm run demo -- output/demo to create and render two variants via the MCP. Use a fresh output folder on each run. The demo uses stock LMMS drums and built-in synthesis.
Run node demo.mjs output/japanese --japanese for a Japanese-inspired take with a D, E-flat, G, A and B-flat pitch palette, more space between plucked notes and sparser kick accents. The demo uses electronic synths and drums. To get closer to acoustic koto or taiko, choose suitable patches or cleared samples in LMMS.
Licence
The bridge is MIT licensed. LMMS is separately distributed under its upstream licence.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Read and edit DAW projects as an agent, save native Logic projects, convert projects between DAWs
Remote MCP server for audio stem separation, denoise, reverb, and MIDI export.
Create, co-edit, analyze, publish, and export collaborative step-sequencer sessions through MCP.
Compose, mix, master and export in Logic Pro, REAPER or Ableton Live by talking to an AI engineer.
Related MCP Servers
- AlicenseBqualityCmaintenanceEnables AI agents to create, modify, and save LMMS music projects programmatically, including tracks, notes, mixer control, effects, and song settings.4524MIT
- AlicenseAqualityBmaintenanceLets a model write FL Studio projects (notes, any installed plugin and preset, mixer, sidechain, automation), open them in a running FL, and render them.157MIT
- AlicenseNot gradedqualityCmaintenanceEnables MCP-capable agents to control a running LMMS session to compose, arrange, mix, and export music, including creating tracks, editing notes, loading presets and samples, automating parameters, and rendering audio or MIDI.GPL 2.0
- AlicenseBqualityCmaintenanceLets MCP clients read and edit a Lemon Studio music project—tempo, notes, channels, patterns, clips, mixer tracks, effects, and plugin parameters—with batched edits that land as a single undo step, plus transport controls for play/pause/stop and project save/render.42MIT