OmniDAW API

The OmniDAW API lets other programs drive Omni: create tracks, write notes and chords, set up effects, automate parameters, listen to the result, master it and export it. The first user is an AI assistant (through the Model Context Protocol, MCP), but the API is plain JSON, so scripts, hardware controllers, test harnesses and your own tools use it the same way.

This section is everything a developer needs to build on it: the concepts and a first call on this page, the wire protocol, the full method reference, a guide to building your own MCP server, client libraries, working recipes, and security and licensing.

How it fits together

 AI client (Claude Desktop, Claude Code, Cursor, ...)
      |  MCP over stdio
      v
 omni-mcp                 the MCP server: tools, prompts and resources
      |                   made from the method catalog
      |  OmniDAW API: JSON-RPC 2.0, one object per line, TCP 127.0.0.1
      v
 Omni                     listener thread -> UI thread -> the same code
      |                   as a click on the screen
      |  engine commands
      v
 audio engine             untouched: the API never runs on the audio thread

Four ideas hold everything together:

1. One catalog. Every capability is a method with a description written for a model, a JSON Schema for its parameters and a ready example. MCP tools, the api.methods document and the reference are all generated from it.
2. The same path as the user. A call does what the matching click does. It runs on the UI thread, so the audio thread is never blocked.
3. Everything reversible. Each edit is one step in Omni's Undo History, labelled API: .... The user takes back anything an assistant did with Ctrl+Z, and the assistant can call edit.undo itself.
4. Ears before hands. Whatever changes the sound can also be measured: meters, loudness, spectrum, stereo width, masking. An agent can check its own work.

What you can do

There are 102 methods in 24 domains. The reference lists all of them.

Turn it on

The API is off by default, because it opens a port on your computer (only reachable from your own machine).

1. In Omni open Settings, AI Create and switch on Enable OmniDAW API. To turn it on for one run only, start Omni with --api or set OMNI_API=1.
2. Omni writes a file with the port and a fresh secret token, endpoint.json, in its settings folder:

The file is api/endpoint.json inside that folder and exists only while Omni runs with the API on. Clients read it to find Omni; nothing else on the network can.

Your first call

The quickest way in is the Python client, one file that needs nothing but a recent Python 3 (download omnidaw.py):

from omnidaw import Omni

with Omni() as omni:
    print(omni.session.info())
    for track in omni.tracks.list():
        print(track["index"], track["name"], track["kind"])

Behind that is one JSON line per call. This is what the client sends over a plain TCP connection to 127.0.0.1:

{"jsonrpc":"2.0","id":2,"method":"session.info"}

and what comes back:

{"jsonrpc":"2.0","id":2,"result":{"api_version":"0.21.0","app_version":"0.1.0","dirty":false,"engine_running":true,"project_name":"My Song","project_path":"C:\\Music\\My Song.omni","return_bus_count":1,"sample_rate":48000,"scene_count":8,"track_count":5}}

(Results grow with new versions; clients should ignore fields they do not know.) Before the first call, a connection must say hello with the token; the protocol page shows the handshake. The Python client does it for you.

Connect an AI assistant

Omni ships an MCP server, omni-mcp, built from the same catalog. It lands next to the Omni executable. In Settings, AI Create, Copy MCP config gives you the entry to paste into your assistant's configuration:

{
  "mcpServers": {
    "omni": { "command": "C:\\path\\to\\omni-mcp.exe" }
  }
}

For Claude Code: claude mcp add omni -- C:\path\to\omni-mcp.exe.

On Linux Omni is an AppImage and omni-mcp is inside it. Run the AppImage with --mcp to start the MCP server (./Omni-x86_64.AppImage --mcp). Copy MCP config then gives you an entry with the path of your AppImage and that flag; for Claude Code: claude mcp add omni -- /path/to/Omni-x86_64.AppImage --mcp.

The server starts even when Omni is not running; tools then answer with a clear error, and the next call connects by itself once Omni is up, also after it restarts. It also offers ready-made prompts (mix review, loudness check, masking check, mastering for a target, arranging a song, composing a loop) and read-only resources.

Want a different server, or one in another language? That is what Building an MCP server is for.

Versions

The API is versioned major.minor.patch. A client built for an older minor version keeps working, because new versions only add methods and fields. Omni accepts any client with the same major version and a version not newer than its own. Current version: 0.21.

Where to go next

• Just want to try it? Run the recipes: a chord loop with a melody, a filter sweep, a mastering loop, chopping a break.
• Building an MCP server? Read Building an MCP server, which includes a complete working server in about 100 lines.
• Writing a client in another language? The protocol is small: sockets and JSON.
• Looking for a method? The reference has every parameter and a ready request for each.