We wanted to know how far an AI agent can get toward a real melodic house track using only open-source tools. The target was the warm, patient sound of Lane 8: a soft four-on-the-floor kick, pads that breathe in time with it, and builds that take minutes instead of seconds. The tools were Ardour, a free and open-source digital audio workstation, and ardour-mcp, a bridge that lets an AI agent like Claude talk to Ardour directly.
This post covers what Ardour is, what we built, and how you can do the same thing yourself, step by step. It also covers what still needs a human ear before any of this is ready to release.
What Ardour is
A digital audio workstation, or DAW, is the program where music gets made. You record into it, arrange clips on a timeline, mix the tracks together, and export a finished song. Ableton Live, Logic Pro and Pro Tools are DAWs. Ardour is one too, but its source code is open under the GPL license, and it runs on Linux, macOS and Windows.
Ardour was created by Paul Davis. Before he turned to audio software in 1998, he was the second employee at Amazon. He also wrote JACK, the audio connection system that a lot of Linux audio still runs on. Ardour's first release came out in September 2005.
The project has had serious industry backing along the way. Harrison, the company that makes large mixing consoles for film and broadcast, has worked with Paul Davis since 2004, and it builds its commercial Mixbus DAW on top of Ardour's engine. Solid State Logic and the SAE Institute audio school have both sponsored development at different points. Sales of Mixbus help fund the project today.
What Ardour can do
Ardour is built like a traditional studio. It records as many audio and MIDI tracks as your computer can handle. Its editor handles non-destructive cutting, crossfades and comping. Its mixer has a channel strip for every track, with plugin inserts, sends, buses and full automation. It hosts plugins in the LV2, VST3, AU and CLAP formats, and it ships with its own set of effects and a basic synth.
Three features matter for this project. Since version 7, Ardour has had clip launching in the style of Ableton's Session view, so you can trigger loops and scenes live. Since version 3.2, it has had a video timeline for scoring and post-production work against picture. And it's scriptable: Lua scripts can run inside the editor, and an Open Sound Control (OSC) server lets outside programs drive the transport and the mixer.
That last point is what makes it a good home for an agent. Professionals use Ardour for multitrack recording, mixing, and sound-to-picture work. Harrison sells its Ardour-based engine to studios as Mixbus. The scripting and OSC layers mean an AI can do the same jobs a mix assistant would.
How the agent connects: MCP
The Model Context Protocol (MCP) is a standard way to give an AI agent a set of tools. ardour-mcp is an open-source MCP server, written in Python, that translates an agent's requests into OSC messages Ardour understands. Its README lists 111 tools across transport, tracks, mixer, recording, automation and metering. You can ask Claude to “set the pad to minus 4 dB and pan the arp 20 percent left” or “what's the loudness of the master bus?” and it does it.
There's one honest limit. ardour-mcp can't write MIDI notes or edit regions yet. Its roadmap lists that as a future phase. So the agent can mix and steer the session, but it can't compose inside it. We got around that by having the agent write the music as code.
What we built: Ocean View Haikus
The track is called Ocean View Haikus. It runs at 122 BPM in F minor and lasts 128 bars, a little over four minutes. The chords loop every eight bars: F minor 9, D flat major 9, A flat add 9, E flat 6/9. These are the kind of minor-key voicings that never quite resolve, which is a big part of why this style feels wistful.
The hook gives the track its name. A bell plays three phrases in each eight-bar loop, five notes, then seven, then five, like the syllables of a haiku. Ocean waves open and close the track.
Claude wrote a Python script that synthesizes every sound from scratch, with no samples. That covers the kick, clap, hats, a rolling offbeat bass, a detuned pad, a sixteenth-note arpeggio, the bell, the waves and a noise riser for the build. The script also handles the arrangement. The pad's filter opens over the first 16 bars. The breakdown drops the kick and bass. The build sweeps everything open with a snare roll. Then comes a 32-bar drop and an outro that closes back down. The pad, bass and arp all duck under each kick, which is the pumping effect at the center of the Lane 8 sound.
The script writes nine stereo stems, MIDI files for every part so you can swap in better synths, a preview mix, and a Lua script that loads it all into Ardour. Then we measured the mix. The groove and drop sit around minus 12 dBFS RMS, and the breakdown and build sit about 11 dB lower, which is the dynamic shape you want.
Step by step: do it yourself
1. Install Ardour. Download it from ardour.org. The free demo build goes silent after about ten minutes of playback, then gives you a few more minutes at a time. Paying any amount from one US dollar unlocks the full build, which is worth it before you start mixing. Building from source is also free.
2. Create a session. Open Ardour, make a new session at 48 kHz, and leave it open.
3. Turn on OSC. In Edit, then Preferences, then Control Surfaces, tick Open Sound Control. Open its settings, keep the port at 3819, and enable feedback.
4. Install ardour-mcp. Clone the repository from GitHub (raibid-entertainment/ardour-mcp), then run uv sync --all-extras inside it. Register it with Claude Code using claude mcp add --transport stdio ardour --scope user -- uv --directory /path/to/ardour-mcp run ardour-mcp, then restart Claude Code so the tools load. The cloned repo also contains its author's own .mcp.json with paths from their machine. Don't approve those entries if Claude Code offers them.
5. Have the agent write the music. Ask your agent to write a Python script, with numpy and scipy, that renders each part to its own WAV file at your session's sample rate. Give it the tempo, key, chords and section map. Ask it to apply one shared gain to all stems so they sum to just under 0 dBFS. In our first render the stems added up to plus 4 dBFS, which would have clipped Ardour's master bus.
6. Import with Lua. Have the script also write a Lua file that sets the tempo, imports every stem as its own track starting at zero, and drops section markers. In Ardour, open Window, then Scripting, paste the file in, and press Run. One warning: test the Lua against your Ardour version. Ardour ships a headless Lua runner that's handy for this. Our first marker code used a constructor that doesn't exist in Ardour 9.8, and we only caught it by testing.
7. Mix with the agent. With OSC on and ardour-mcp loaded, ask the agent to balance the faders, loop sections for you, add sends to a reverb bus, and check the master loudness.
8. Export. Session, then Export, then WAV at 24-bit and 48 kHz.
Can this reach production quality?
Ardour can. It's the engine under a commercial DAW sold to professional studios, and it handles everything from tracking a band to mixing a film. The limits sit in the sounds you feed it and the ears doing the mixing.
Our draft is a sketch. Synthesizing a pad from raw saw waves in numpy gets the shape and movement right, but it doesn't sound as rich as a dedicated synth. To take a track like this toward release:
- swap the MIDI files onto a strong free synth like Surge XT
- replace the code-built drums with well-recorded samples
- mix on headphones or monitors you trust
- master to a streaming loudness target
The agent speeds up the setup, the arrangement and the busywork. Deciding when the track feels right still takes a person listening.
Our own track is at that point now. The stems and the import script exist, and the next step is to bring them into Ardour, switch on OSC and let the agent start mixing while we listen.
Why this matters
Open-source tools and agents fit together well. Ardour is free, so nothing stops you from starting tonight. Its scripting and OSC layers let an agent take over the repetitive work. And because the music lives in a script, you can change one line and re-render the whole track in about a minute. That makes it easy to try a different key, a slower build or a new haiku melody.
Links
Argo Writer Agent