Audionaut

audionaut-mcp

An MCP server that lets AI agents (Claude Code, Claude Desktop, and any other MCP client) edit Audionaut projects: cut, move, fade, analyse, auto-edit and export. It is a thin wrapper: every tool runs one Audionaut command with --json and relays the result — no engine logic lives here.

Requires the Audionaut app (Mac App Store / GitHub Releases). On its own this package does nothing.

Quick start

You need Audionaut — 1.6.3 or later is recommended; 1.6.2 works on macOS and Linux — and Node.js 18 or later.

Claude Code:

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

Claude Desktop (claude_desktop_config.json):

{

"mcpServers": { "audionaut": { "command": "npx", "args": ["-y", "audionaut-mcp"] } } }

On Windows, if Claude Desktop cannot start npx, use "command": "cmd", "args": ["/c", "npx", "-y", "audionaut-mcp"].

Then ask for something like *"split the clip in ~/Music/Audionaut/Demo Project.audium every 16 bars"*. Keep the project open in Audionaut to watch the edits arrive; each one is a single undo step.

On macOS, keep projects in your Music folder. Audionaut is sandboxed there (App Store and website download alike) and can only reach ~/Music: projects, audio to import and export targets all have to be inside it, for example in ~/Music/Audionaut. The server refuses other paths up front with sandbox_denied and a hint. Windows and Linux have no such limit.

To see which Audionaut the server found and whether it answers:

npx -y audionaut-mcp --check

The server looks for, in order: AUDIONAUT_CLI; a developer build of audionaut-cli (see the end of this page); audionaut-cli on PATH; the installed app — /Applications or ~/Applications (or wherever Spotlight finds it) on macOS, Program Files\Audionaut on Windows, audionaut from the .deb or an Audionaut*.AppImage in ~/Applications or ~/.local/bin on Linux. Anywhere else, point AUDIONAUT_CLI at the binary, e.g. your AppImage.

separate_stems needs the Demucs model, which you download once in the app (Settings ▸ Separation). The model weights are licensed for research use.

A stem split takes minutes for a full song, and exporting or analysing a long project can too. When your client sends a progressToken with the call, separate_stems, export_audio and analyze report progress (Audionaut 1.6.5 and later). A call is only stopped after ten minutes without any sign of life, not ten minutes in total.

Projects that are open in Audionaut

The tools do not need the user to save first, and they will not write over a document the user has open. When Audionaut is holding the project, the CLI hands the command to the app: it acts on the document as the user currently sees it, unsaved changes included, and the result lands in their undo history as one step instead of touching Project.json. Saving stays theirs to do.

With no app holding the project, the tools read and write the project file as they always have. If an app is holding one but cannot be reached, the command fails (host_unavailable) rather than falling back to the file, which would be missing whatever is unsaved.

Autosave.json and Host.json inside a package belong to the app — never read or edit them.

Tools

A typical agent flow: create_project → import_audio → analyze → auto_edit/assemble → export_audio. CLI errors come back as tool errors carrying the CLI's own code: message (e.g. demucs_unavailable: ... in builds without stem separation), so agents can react.

Feature requests and bug reports from agents

Agents run into the edges of what the tools expose, and into their bugs, long before a person would file an issue, so two tools give them a direct channel. request_feature takes a title, a description and (ideally) what the agent was trying to do and which tool fell short. report_bug takes the same plus steps to reproduce, the expected behaviour and the project shape; its description tells agents to quote the exact tool call and error text and never to attach audio. The server's instructions point agents at the right one whenever a task needs something the tools cannot do or a tool misbehaves, and tell them to say so to the user. A reporter name or e-mail is included only when the user offers one. Both become GitHub issues - enhancement + agent-request for features, bug + agent-report for bugs; the reply carries the issue URL and the feature-request discussion.

Transports, tried in order:

1. Relay — the default: a deployment of Tools/feature-request-relay, a Cloudflare Worker that files the issue with its own GitHub token. This is the path for end users, who have no GitHub credentials on the agent side. AUDIONAUT_FEATURE_REQUEST_URL points at another deployment; an empty value skips the relay.
2. GitHub CLI — when the relay is skipped and gh is installed and logged in, the issue is filed with gh issue create under the user's own account (developer machines).
3. Nothing — otherwise the reply says sent: false and carries a prefilled new-issue URL for the user to open themselves.

AUDIONAUT_DISABLE_FEATURE_REQUESTS=1 skips the first two for both tools (test harnesses, air-gapped setups); the smoke test points the relay URL at a local stand-in.

Environment variables

The server sets two variables for every Audionaut it runs, which the open app uses to group an agent's edits into undo steps: AUDIONAUT_AGENT_SESSION (one id per server process) and AUDIONAUT_AGENT_TASK (the editing tool's task parameter). Edits with the same session and task share one undo step in the user's app, and edits that name no task share one step per session; a user edit, an undo or ten idle minutes start a new one.

Developing: building the CLI from source

From a clone of the repo, the server prefers a CMake build of the standalone audionaut-cli, which is not sandboxed:

1. Build the CLI (from the repo root):

   cmake -B build -S Audionaut/Catch2Tests

cmake --build build -j8 --target AudionautCli

1. Install the server's dependencies:

   cd Tools/audionaut-mcp && npm install

1. Register the checkout instead of the npm package. Claude Code:

   claude mcp add audionaut -- node /path/to/Audionaut/Tools/audionaut-mcp/index.js

Claude Desktop: the JSON above with "command": "node" and "args": ["/path/to/Audionaut/Tools/audionaut-mcp/index.js"].

npm test drives the server through a real MCP client against whichever binary it finds (AUDIONAUT_CLI picks one). With the sandboxed macOS app it works in a scratch folder under ~/Music; AUDIONAUT_TEST_DIR overrides the location.

Paths in tool arguments are best given absolute — relative paths resolve against the server process's working directory, which depends on the MCP client.