LOCAL AUDIO WORKFLOWS · CHECKED OCTOBER 6, 2026

How to Separate Audio Stems with an AI Agent

Yes: an MCP-capable agent can ask Audionaut to separate a recording locally. The agent controls the workflow; HTDemucs does the separation. You need the desktop app, its downloaded model and an MCP connection—not just a chat window.

This guide covers Audionaut's documented workflow. It is a source-checked setup guide, not a hands-on audio benchmark. Work only with recordings you own or have permission to process.

What is Audionaut?

Audionaut is an open-source multitrack audio editor and recorder. Its project repository links the desktop application and agent tooling. It suits people who want separation inside an editable project, rather than a standalone upload-and-download task.

What actually separates the audio?

The official stem separation manual documents local HTDemucs through demucs.cpp. One selected mono or stereo clip produces four aligned tracks: drums, bass, other and vocals. The documented clip limit is ten minutes. Download the model once in Settings → Separation before asking an agent to use it.

For a manual check, stop playback, select one clip and use Edit → Separate Stems. Listen to the resulting tracks against the source. Separation estimates parts from a mixture; it does not recover guaranteed clean studio masters.

Connect an MCP client

The maintainer's MCP README recommends Audionaut 1.6.3 or later and Node.js 18 or later. With the app installed, Claude Code can register the server using:

claude mcp add audionaut -- npx -y audionaut-mcp

Then check that the server can find and contact the app:

npx -y audionaut-mcp --check

Claude Desktop has a different configuration entry; follow the same README's client-specific instructions. On macOS, put projects, imported audio and export destinations inside ~/Music for the sandboxed app. Use explicit project and track references rather than asking the agent to guess which recording you meant.

From a request to four tracks

1. Claude or another MCP client receives your instruction.
2. Audionaut MCP calls separate_stems for the project and track, optionally identifying a clip.
3. The local HTDemucs separator processes that clip.
4. Audionaut reports the new drums, bass, other and vocals tracks.
5. You review them, then choose the edits and export you need.

For example, ask: “Separate the selected recording into stems. Let me review the vocals before making an instrumental export.” This is an illustrative instruction, not a tested promise that an agent can finish every edit correctly. Specify the desired change and inspect the output before using it.

CLI, errors and privacy boundaries

The official CLI guide lists separate as the corresponding command. Consult your installed CLI help for arguments; the MCP tool name is not itself a shell command. Commands return structured results so the client can distinguish a completed edit from an error.

If the app cannot be found, resolve the server check first. If separation cannot start, check the model download and whether that build includes Demucs. For a macOS sandbox_denied error, move the project and audio into the documented Music folder. Do not keep retrying an export without checking where it will write.

Local separation does not upload the recording to a web separator. Your chosen agent service may still receive conversation text and tool results; review its data settings separately. The MCP README also labels the bundled separation weights as research-use weights. Check the applicable model terms before choosing a production use.

Local agent workflow or a browser?

Choose Audionaut when local processing and project automation justify the setup. Choose a browser workflow when you want a straightforward separation without installing a local model. This is a workflow comparison; we have not run comparable speed, price or quality benchmarks.

Questions before you start

Can Claude itself separate vocals?

In this workflow Claude selects and calls a tool. The separator is running in Audionaut, not inside the language model. A conversation without the connected tool cannot perform this local operation.

Do I need a local GPU for Extract Stems?

No local model installation is needed for its browser workflow. Your authorized audio is uploaded for processing. If keeping the recording on your Mac is the priority, see the Apple Silicon local guide.