Open-source editor Audionaut lets AI agents edit multitrack audio sessions over MCP

A multitrack editor built for humans and agents alike

Audionaut, a free and open-source multitrack audio editor that reached the Hacker News front page on 2 October 2026, is built around an unusual premise: it is meant to be driven by AI agents as well as people. The desktop app exposes its editing engine over the Model Context Protocol (MCP), so tools such as Claude can cut, arrange and export sessions directly.

According to the project's GitHub repository, Audionaut is written in C++ on the JUCE framework and runs natively on Windows, macOS and Linux. It is aimed at music, podcasts and general multitrack recording — fine-grained cutting, per-track playlists, multi-channel handling and straightforward exports — while stopping short of a full digital audio workstation.

One-line MCP setup, one undo step per edit

The agent integration is deliberately lightweight. With the app running and Node.js 18 or newer installed, the README shows a single command to register the MCP server with Claude: claude mcp add audionaut -- npx -y audionaut-mcp. Edits made by the agent then land in the open project, and each one registers as a single undo step, so a human can step through or revert what the agent did. On macOS, the documentation advises keeping projects inside the Music folder.

DSP extras: Essentia analysis and Demucs stem separation

Two heavier audio features ship as submodules. Analysis — BIC segmentation, onset detection and beat tracking — links against a static build of Essentia. Building it is optional: the app and its tests compile with analysis features disabled when Essentia is absent. The build path has sharp edges. Essentia's bundled waf needs distutils, so Python 3.11 or older is required, and CMake 3.x is recommended because some vendored build files are too dated for CMake 4.x to accept.

Stem separation comes from demucs.cpp, a C++ port of Meta's Demucs, compiled directly from a submodule with no separate build step. It splits a track into drums, bass, other and vocals. The model weights are not stored in the repository; the app downloads them on first use into its Models folder, or the CLI can be pointed at a local copy.

A headless CLI underneath the MCP server

For scripts, CI and agents that do not speak MCP, the project also builds audionaut-cli, which manipulates .audium project files without a GUI or audio device. Every command accepts -- and prints exactly one machine-readable envelope on stdout — an ok/result or ok/error pair — with logging pushed to stderr. Exit codes are documented: 0 for success, 1 for a failed operation, 2 for a usage error and 3 when a feature is unavailable in that build, such as analyze without Essentia.

The verb set covers a complete workflow: create, import, analyze, auto-edit, assemble and split, plus region and clip operations like create-region, place-clip, move-clip, clip-gain and clip-fades, ending in export with explicit sample-rate and bit-depth options. The README sketches a typical agent flow — create, import, analyze, auto-edit or assemble, then export — checking the ok flag in each response along the way. Projects round-trip cleanly: files written by the CLI open in the GUI app and vice versa.

The GUI binary accepts the same verbs and executes them headlessly, exiting with the command's exit code even while a GUI instance is open. There are platform caveats. The sandboxed macOS app can only reach entitled locations such as ~/Music from its in-app CLI, so the standalone build is needed for wider file access. On Windows the app is a GUI program that returns the shell prompt immediately, making audionaut-cli the better scripting option there.

CLI invocations report anonymous usage statistics — one event recording the verb and exit code — but only under the same strictly opt-in consent used by the app, and setting AUDIONAUT_DISABLE_ANALYTICS=1 switches reporting off regardless, which matters for CI environments.

License and building

Audionaut is dual-licensed under GPLv3-or-later and a commercial license. Contributions require a signed contributor license agreement, and development is funded partly through sponsorship. Building from source requires cloning with submodules, and the repository ships an Xcode project, a Visual Studio 2026 solution and a Linux Makefile target.

Why it matters

Most audio editors are GUI-first, which leaves automation to brittle scripting hooks or screen control. Audionaut treats agents as first-class users: MCP for interactive editing, a JSON-emitting CLI for pipelines, and undo steps and exit codes that make agent actions inspectable and reversible rather than opaque. The feature set also sketches a plausible automated workflow — import takes, analyze them, auto-edit into a structure, export a mix — that podcast and rough-mix production could adopt quickly. The caveats are real: the details come from the project's own README as posted to Hacker News rather than independent testing, the Essentia build is fussy about Python and CMake versions, and stem separation depends on a model downloaded on first run. Still, as a signal of where desktop software is heading — native, agent-ready interfaces rather than bolted-on automation — it is a project worth watching.