stemsplit-mcp

AI stem separation and voice cleaning as a Model Context Protocol (MCP) server. Remove vocals, build karaoke tracks, isolate dialogue, split any song into vocals, drums, bass, piano, guitar, and other stems — or remove background noise from voice recordings using DeepFilterNet — directly from Claude Desktop, Cursor, Cline, Windsurf, Zed, or any other MCP-compatible client. Works with local audio files (MP3, WAV, FLAC, M4A, OGG, AAC) and YouTube/SoundCloud URLs.

Powered by the StemSplit API (HTDemucs for stem separation, DeepFilterNet for noise removal). The server exchanges only file paths and JSON over MCP — audio bytes never pass through the LLM context. They flow directly between your machine, StemSplit's API, and Cloudflare R2.

What you can do with this

Audio separation basics

• Remove vocals from a song — separate any MP3, WAV, or FLAC into vocals and instrumental

• Build a karaoke version of any track — /karaoke slash command returns just the instrumental

• Extract an acapella — pull a clean vocal track for remixes, mashups, or re-arrangement

• Extract drums, bass, piano, or guitar — split audio into up to six individual stems

• Process YouTube videos — paste a youtube.com or youtu.be URL and get separated stems back

Audio production & post-production

• Clean vocals before processing — isolate vocals first, then pass to a de-esser, noise reducer, or pitch corrector without mix bleed affecting the result

• Stem delivery for mastering — auto-generate per-stem exports from a final mix for a mastering engineer

• Adaptive game audio — split a track so a game engine can fade individual layers (e.g. mute drums during quiet scenes)

• DJ acapella/instrumental packs — batch-generate acapellas and instrumentals for live performance or DJ sets

• Sample chopping — extract drums or bass for sample packs in hip-hop / electronic production

Voice cleaning & noise removal

• Clean up a podcast or interview — remove hum, hiss, HVAC noise, or ambient room sound from any voice recording

• Denoise vocals after stem separation — pass denoiseVocals: true to separate_stems and get a noise-free vocals stem in one shot

• Clean dialogue for video production — strip wind, echo, or background noise before syncing to picture

• Pre-process audio before transcription — clean first for dramatically higher ASR / Whisper accuracy

AI & developer pipelines

• Vocals → transcription — isolate vocals first, then feed to Whisper or any ASR model for significantly cleaner speech-to-text

• Lyrics generation — vocals → transcription → synced lyrics file, fully automated in a single MCP chain

• Training data for AI music models — generate clean separated stems from raw mixed tracks for fine-tuning or dataset building

• Content-ID / copyright checking — extract vocals to fingerprint and match against a vocal database

• Per-stem audio visualizers — drive instrument-reactive visualizers in video or web apps by separating stems first

Content & media

• Podcast / interview cleanup — strip music beds or background music from recorded dialogue

• Sync licensing — instantly generate an instrumental version of a submitted track for a music supervisor

• Music education apps — isolate individual instruments to build solo/mute practice tools or ear training exercises

Agentic workflows

• Build audio agents in your IDE — orchestrate stem separation from Cursor or Claude Desktop using natural language

• Batch process audio in MCP-driven pipelines — chain stem separation with transcription, translation, or any other MCP tool

Related MCP server: MusicGPT MCP Server

MCP clients supported

stemsplit-mcp runs as a local stdio MCP server, so it works in any client that supports the standard MCP transport:

• Claude Desktop (Anthropic)

• Cursor

• Cline (VS Code extension)

• Windsurf (Codeium)

• Zed

• Any client following the Model Context Protocol specification

Tools, resources, and prompts

Stem separation

Voice Cleaner (noise removal)

Account

Plus six ready-made prompts (slash commands): karaoke, isolate_dialogue, sampler_pack, youtube_instrumental, soundcloud_instrumental, clean_voice.

Install

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "stemsplit": {
      "command": "npx",
      "args": ["-y", "stemsplit-mcp"],
      "env": {
        "STEMSPLIT_API_KEY": "sk_live_your_key_here"
      }
    }
  }
}

Restart Claude Desktop. Type /karaoke or just ask: "Separate the vocals from ~/Music/demo.mp3".

Cursor

Add to ~/.cursor/mcp.json (or per-workspace <workspace>/.cursor/mcp.json):

{
  "mcpServers": {
    "stemsplit": {
      "command": "npx",
      "args": ["-y", "stemsplit-mcp"],
      "env": {
        "STEMSPLIT_API_KEY": "sk_live_your_key_here"
      }
    }
  }
}

Cline, Windsurf, Zed, others

Any MCP client that supports stdio-launched servers works. Use the same npx -y stemsplit-mcp command and pass STEMSPLIT_API_KEY via the client's env mechanism.

Get an API key

1. Sign up at stemsplit.io

2. Open stemsplit.io/app/settings/api

3. Generate a key (format: sk_live_...)

4. Paste it into your MCP client config as shown above

Configuration

Tool reference

separate_stems

Submit an audio file or direct URL for stem separation.

{
  "source": "/Users/me/Music/song.mp3",
  "outputType": "BOTH",
  "quality": "BEST",
  "outputFormat": "MP3",
  "wait": true
}

Returns (wait=true):

{
  "jobId": "job_abc123",
  "status": "COMPLETED",
  "creditsCharged": 180,
  "outputDir": "/Users/me/Downloads/stemsplit/job_abc123",
  "stems": {
    "vocals": "/Users/me/Downloads/stemsplit/job_abc123/vocals.mp3",
    "instrumental": "/Users/me/Downloads/stemsplit/job_abc123/instrumental.mp3"
  }
}

separate_youtube

Same shape, but takes youtubeUrl instead of source. Output is fixed to vocals + instrumental, MP3, BEST quality (this is the StemSplit API's contract for YouTube jobs).

clean_voice

Submit an audio file or direct URL for noise removal using DeepFilterNet. Removes background hum, hiss, HVAC noise, wind, echo, and other ambient sounds. By default (wait=true), polls until complete and downloads the cleaned audio to disk.

{
  "source": "/Users/me/recordings/podcast-ep12.mp3",
  "outputFormat": "MP3",
  "wait": true
}

Returns (wait=true):

{
  "jobId": "dnz_abc123",
  "status": "COMPLETED",
  "creditsCharged": 180,
  "outputDir": "/Users/me/Downloads/stemsplit/dnz_abc123",
  "cleanedAudioPath": "/Users/me/Downloads/stemsplit/dnz_abc123/podcast-ep12_denoised.mp3"
}

get_job, list_jobs, get_youtube_job, list_youtube_jobs, get_denoise_job, list_denoise_jobs, get_balance, download_stems

Thin wrappers over the corresponding StemSplit /api/v1 endpoints. download_stems re-fetches the job first to mint fresh 1-hour presigned URLs, so the expiry never matters. get_denoise_job returns outputs.audio.url when the job is COMPLETED.

Resources

Read-only context the LLM can pull on demand.

Prompts (slash commands)

Example sessions

Karaoke from a local file (Claude Desktop):

Make a karaoke version of ~/Music/demo.mp3.

Claude calls separate_stems with outputType="BOTH", polls for ~60s, and returns:

Done. Karaoke (instrumental) is at:
/Users/me/Downloads/stemsplit/job_abc123/instrumental.mp3

Six-stem sampler pack (Cursor):

Split ./loops/break.wav into all six stems for sampling.

Cursor calls separate_stems with outputType="SIX_STEMS", quality="BEST", outputDir="./loops/break-stems", and reports each file path so you can drop them into your DAW.

Instrumental from YouTube:

Get me the instrumental of https://youtu.be/dQw4w9WgXcQ.

Claude calls separate_youtube, polls until COMPLETED, downloads vocals.mp3 and instrumental.mp3 to ~/Downloads/stemsplit/<jobId>/, and returns the instrumental path.

Clean vocals for transcription (Claude Desktop):

Transcribe the lyrics from ~/Music/interview-with-music.mp3 — there's a music bed underneath, clean it up first.

Claude calls separate_stems with outputType="VOCALS" to strip the music bed, then passes vocals.mp3 to a transcription tool (e.g. Whisper via another MCP server). The result is a clean transcript with none of the background music interfering.

Batch acapella extraction (Cursor agent):

Extract acapellas from every MP3 in ./tracks/ and save them to ./acapellas/.

Cursor iterates the directory, calls separate_stems with outputType="VOCALS" and a custom outputDir per file, and returns a list of acapella paths ready for a remix session or AI training dataset.

Vocal isolation → stems for remix (Claude Desktop):

I want to remix ~/Music/original.wav. Give me the acapella and all the individual instrument stems separately.

Claude calls separate_stems twice — once with outputType="VOCALS" for the clean acapella, once with outputType="SIX_STEMS" for the full stem pack — and hands back all seven file paths organized by stem type.

Clean up a noisy podcast recording (Claude Desktop):

My recording at ~/Desktop/episode-23.mp3 has a lot of HVAC noise. Can you clean it up?

Claude calls clean_voice with source="~/Desktop/episode-23.mp3", polls for ~10–20s, and returns:

Done. Cleaned audio saved to:
/Users/me/Downloads/stemsplit/dnz_abc123/episode-23_denoised.mp3

Separate stems and denoise vocals in one call (Cursor):

Extract the vocals from ./stems/song.wav and clean them up for use in a transcription pipeline.

Cursor calls separate_stems with outputType="VOCALS" and denoiseVocals=true. StemSplit runs Demucs to extract the vocals, then immediately passes them through DeepFilterNet — delivering a single noise-free vocals file in one job.

Voice Cleaner (noise removal)

Use clean_voice (or the /clean_voice slash command) to remove background noise from any audio file using StemSplit Voice Cleaner, powered by DeepFilterNet.

{
  "source": "/Users/me/recordings/episode.mp3",
  "outputFormat": "MP3"
}

What it removes: background hum, HVAC/air conditioning noise, hiss, wind, echo, ambient room sound, and other non-speech noise. Ideal for podcasts, voiceovers, interview recordings, and dialogue.

Limits

• Maximum duration: 60 minutes

• Output: a single cleaned audio file (same duration as input)

• Credits: 1 credit = 1 second of audio

Example (Claude Desktop)

Clean up the background noise in ~/Recordings/interview.wav.

Claude calls clean_voice, polls until COMPLETED (~10–30s depending on length), and returns:

Done. Cleaned audio saved to ~/Downloads/stemsplit/<jobId>/interview_denoised.wav

YouTube stem separation

Use separate_youtube (or the /youtube_instrumental slash command) to extract vocals and an instrumental from any YouTube video.

{
  "youtubeUrl": "https://www.youtube.com/watch?v=dQw4w9WgXcQ"
}

Accepted URL formats

Limits

• Maximum duration: 60 minutes

• Output: vocals + instrumental, MP3, BEST quality (fixed)

• Credits: 1 credit = 1 second of video

Example (Claude Desktop)

Get me the instrumental of https://youtu.be/dQw4w9WgXcQ.

Claude calls separate_youtube, polls until COMPLETED (~60s for a 3-minute video), and returns:

Done. Files saved to ~/Downloads/stemsplit/<jobId>/
  vocals.mp3
  instrumental.mp3

SoundCloud stem separation

Use separate_soundcloud (or the /soundcloud_instrumental slash command) to extract vocals and an instrumental from any public SoundCloud track.

{
  "soundcloudUrl": "https://soundcloud.com/artist/track-name"
}

Accepted URL formats

Limits

• Maximum duration: 15 minutes

• Must be a public track (private tracks and sets/playlists are not supported)

• Output: vocals + instrumental, MP3, BEST quality (fixed)

• Credits: 1 credit = 1 second of audio. When track duration is unknown at submission, 4 minutes (240 credits) is held and reconciled on completion.

Example (Claude Desktop)

Remove the vocals from https://soundcloud.com/artist/my-track.

Claude calls separate_soundcloud, polls until COMPLETED, and returns:

Done. Files saved to ~/Downloads/stemsplit/<jobId>/
  vocals.mp3
  instrumental.mp3

Example (Cursor agent)

Extract the acapella from every SoundCloud URL in ./tracks.txt and save each to ./acapellas/.

Cursor reads the file, iterates the URLs, calls separate_soundcloud with outputDir set per track, and returns a list of all saved acapella paths.

Supported inputs

• Local files: mp3, wav, flac, m4a, ogg, webm, aac, wma

• Direct URLs: any public https:// URL serving one of the formats above (the StemSplit API fetches it server-side)

• YouTube: youtube.com/watch?v=..., youtu.be/..., youtube-nocookie.com/embed/..., or a bare 11-character video ID

• SoundCloud: soundcloud.com/artist/track, m.soundcloud.com/artist/track, or on.soundcloud.com/shortcode (public tracks only, max 15 minutes)

Limits: 100 MB / 60 minutes per file. 1 credit = 1 second of audio. Credits are deducted at job submission.

Troubleshooting

Development

git clone https://github.com/StemSplit/stemsplit-mcp
cd stemsplit-mcp
npm install
npm run typecheck
npm run lint
npm test
npm run build

STEMSPLIT_API_KEY=sk_live_... npm run inspect

npm run inspect launches the MCP Inspector for interactive testing.

FAQ

How do I remove vocals from a song in Claude Desktop?

Add the install snippet above to claude_desktop_config.json, restart Claude, then ask:

Remove the vocals from ~/Music/song.mp3.

Claude calls the separate_stems tool, waits for the job to complete (~30–60s for a 3-minute track), and hands back the local path to the instrumental file. Or use the /karaoke slash command directly.

Can this work with YouTube URLs?

Yes. Use the separate_youtube tool or the /youtube_instrumental slash command. The StemSplit API handles the YouTube download server-side and returns vocals + instrumental stems. Output is fixed to vocals + instrumental, MP3, BEST quality.

What stems can I extract?

Vocals, instrumental, drums, bass, other, piano, and guitar. Six-stem output (adding piano and guitar) requires quality=BEST and is only available for stem jobs (not YouTube jobs).

How is this different from the StemSplit web app?

The web app is point-and-click. This MCP server lets you orchestrate stem separation through natural-language prompts to an LLM, or programmatic tool calls from any MCP client. Same backend (HTDemucs / Demucs on GPU), different interface. Use the web app for one-off jobs; use the MCP server when you want to chain stem separation with other tools (transcription, translation, agentic pipelines) inside an LLM-driven workflow.

Does this run the AI model locally?

No. The MCP server is a local stdio process that talks to the StemSplit cloud API over HTTPS. Audio bytes are uploaded directly to Cloudflare R2 via presigned PUT (your API key never crosses the network with the audio). Stem separation runs on StemSplit's GPU workers. If you want fully local separation, look at demucs or demucs-onnx.

How much does it cost?

StemSplit uses a pay-per-second model: 1 credit = 1 second of audio. Credits are deducted at job submission. New accounts include free credits. Check current pricing at stemsplit.io/pricing.

What audio formats are supported?

Input: MP3, WAV, FLAC, M4A, OGG, WebM, AAC, WMA (up to 100 MB / 60 minutes). Output: MP3, WAV, or FLAC.

Where do the stems end up?

By default, in ~/Downloads/stemsplit/<jobId>/ with one file per stem. Override per-call with outputDir or globally with the STEMSPLIT_DEFAULT_OUTPUT_DIR env var.

Can I use this in a custom MCP client or LangChain agent?

Yes. stemsplit-mcp follows the MCP spec exactly. Any client that speaks the stdio transport works. For programmatic Node.js / TypeScript clients, see @modelcontextprotocol/sdk.

Can I remove background noise from a recording?

Yes. Use the clean_voice tool (or the /clean_voice prompt). It runs DeepFilterNet on your audio and returns the cleaned file. You can also pass denoiseVocals: true to separate_stems to denoise the extracted vocals stem automatically as part of a stem separation job.

What if the job takes longer than the timeout?

Pass wait: false to separate_stems, separate_youtube, or clean_voice. You'll get the jobId back immediately and can poll later with get_job / get_youtube_job / get_denoise_job. Or set a longer timeoutSeconds (up to 3600s).

How do I get an API key?

Sign up at stemsplit.io and generate a key at stemsplit.io/app/settings/api. The key format is sk_live_....

License

MIT (c) 2026 StemSplit

Related projects

• StemSplit — hosted stem separation web app and API

• StemSplit API docs — full REST reference + OpenAPI spec

• n8n-nodes-stemsplit — n8n community node for stem separation workflows

• stemsplit-python — Python SDK for the StemSplit API

• stemsplit CLI — command-line tool (Go), available via Homebrew

• demucs-onnx — ONNX export of HTDemucs for local inference

• Model Context Protocol — the open protocol this server implements

Keywords

stem separation MCP, vocal remover MCP, karaoke generator MCP, voice cleaner MCP, noise removal MCP, background noise remover, DeepFilterNet MCP, Claude Desktop audio, Cursor audio tools, instrumental extractor, acapella extractor, AI stem splitter, MCP audio server, remove vocals from MP3, isolate vocals, split audio into stems, YouTube vocal remover, SoundCloud vocal remover, SoundCloud stem separator, SoundCloud instrumental extractor, HTDemucs MCP, Demucs MCP, MCP server for stem separation, podcast noise removal, audio cleanup AI.