October 31, 2026 · Pirkka Räisänen

How to Control a DAW with AI Agents: The Songen MCP Server

Give your AI agents the ability to read timelines, generate MIDI, automate plugins, and render audio stems directly in Songen using the Model Context Protocol (MCP).

You control Songen from an MCP-capable AI client by connecting it to Songen’s Model Context Protocol (MCP) server on Mac. The model reads your open project, sends structured edits, and can render short audio passes to verify level and tone. You still play the app yourself; the agent handles repetitive arrangement and mix checks.

The MCP integration allows chat models to interact directly with your open session. Your AI agent can move a chorus, nudge bus EQ, and edit the timeline you are actively working on, all without sending your audio to a cloud renderer.

Songen for Mac now ships the server built in. Point an AI client (for example Claude Desktop, Cursor, or Codex) at it and they talk to the Vibin AI Audio Engine through a local bridge.

What is the Songen MCP Server?

MCP (Model Context Protocol) is an open wire format for hooking an AI client to local tools and files.

On your Mac, Songen listens on 127.0.0.1 only. It serializes the open song (sections, clips, notes, mixer, automation) for the client. You skip fake UI clicks. The client:

  1. Reads the arrangement.
  2. Proposes edits as patches.
  3. Applies them to the project file (Undo still works).
  4. Renders and analyzes offline audio when it needs ears on the result.

Rendering stays on the machine. The LLM only sees what you paste or what the client sends upstream.

What can AI agents do in Songen?

You get faster passes on layout, automation, and balance. The engine schema is strict, so multi-step plans fail early instead of mangling the file.

Build and reshape arrangements

Start from one 8-bar loop. Ask the agent to duplicate sections, swap which tracks play in the verse, or stretch the form to 64 bars without clicking every cell.

Write automation and movement

Name the move: filter sweep on the build, wider chorus pan, bass ducked under the kick. Songen can reset parameters at section boundaries so the next block starts clean.

Mix and balance with offline renders

The client requests a bar-range bounce, reads spectrum or stem balance, then patches gain or EQ. You approve applies; renders stay local.

Example prompts for MCP clients

Copy these, or grab more on the MCP setup guide:

“Read my project state and schema. I have one 8-bar loop on the timeline. Propose a 64-bar arrangement: intro 8, verse 16, chorus 8, verse 16, chorus 8, outro 8. Validate then apply; render bars 1–8 and 25–32 to confirm.”

“Find the master EQ low-pass cutoff control. Automate 1400 Hz on all verse columns and full bandwidth on choruses with restore at boundaries. Validate with spectrum on verse vs chorus.”

“Analyze mix balance on the chorus. Bring bass up 2 dB and reduce lead 1 dB if stems are unbalanced; verify with analyze_mix_balance after apply.”

Security and edit safety

  • Localhost only: Nothing listens on your LAN. Songen keeps a revocable token in Keychain.
  • Validate, then apply: Run validate_project_patch on the current revision before apply_project_patch.
  • Toggle: Connection works when Songen is open and Let AI apps edit Songen is on in Preferences.

Licensing the engine

The MCP server sits on the Vibin AI Audio Engine, the same graph Songen uses for generative MIDI and offline bounces. Platforms building their own AI music stack can license the engine.

Download Songen for Mac, then walk through the MCP setup guide to link your client.