The tools wait for the job and download the stems into a folder you chose, so the result is something you can open in a DAW rather than a link to chase. Audio bytes travel between your machine and StemSplit directly — they are never carried through the conversation.
StemSplit's own MCP server, run through npx. It takes a local file, a direct audio URL, a YouTube link or a SoundCloud track, sends it to the StemSplit API — HTDemucs for separation, DeepFilterNet for noise removal — polls the job, and writes the audio to a directory you name. Only paths and JSON cross MCP; the audio itself never enters the model's context.
- Vocals, instrumental, both, four stems or six — the six-stem split adds piano and guitar and runs at BEST quality
- YouTube and SoundCloud links handled server-side, returning vocals plus instrumental at BEST quality in MP3
- Background hum, HVAC noise, hiss, wind, echo and room tone stripped from a voice recording, up to 60 minutes long
- Separation and denoise in a single job — pass denoiseVocals: true and the vocals stem comes back already cleaned
- Output as MP3, WAV or FLAC, written under ~/Downloads/stemsplit/<jobId>/ unless you pass outputDir
- Job status and history for every kind of job, plus the remaining credit balance
- A completed job re-downloadable later: download_stems re-fetches it and mints fresh 1-hour presigned URLs
- Six slash-command prompts for the usual jobs — karaoke, isolate_dialogue, sampler_pack, youtube_instrumental, soundcloud_instrumental and clean_voice
Node.js 20 or newer, a StemSplit account, and a key created in settings, passed as STEMSPLIT_API_KEY — it must start with sk_live_. Runs over stdio from npm as stemsplit-mcp. Two optional variables move the endpoint and the output directory: STEMSPLIT_API_BASE_URL and STEMSPLIT_DEFAULT_OUTPUT_DIR. Credits are charged by length, one credit per second of audio.
One command plus a key — npx -y stemsplit-mcp, then supply credentials