Logic Pro MCP

Logic Pro MCP Server for Claude, Cursor, and AI Agents

A local Model Context Protocol (MCP) server that lets Claude Code, Claude Desktop, Cursor, Codex, VS Code, and other MCP clients control Logic Pro for AI music production: create tracks, write MIDI, operate transport and mixer state, inspect live project data, and verify results.

Install · Watch demo · What it controls

### 🏆 Selected for Anthropic's Claude for Open Source program

Logic Pro MCP has been officially selected for Anthropic's Claude for Open Source program — recognition from the makers of Claude that this project is open-source work worth supporting. This server is built with Claude, for Claude-powered agents — and is now officially supported by the program.

An 82 BPM D-minor lofi loop composed live in Logic Pro 12.3 by the MCP — tempo, three MIDI parts, a Drummer, and playback, all from prompts.

▶ Watch the 36-sec demo (with sound)

---

Logic Pro MCP Server gives Claude, Cursor, Codex, and other MCP clients a structured way to control Logic Pro without brittle keyboard macros. Logic Pro does not ship a first-party API for agentic composition, session setup, mixer operations, or live project readback, so Logic Pro MCP fills that gap by combining 7 native macOS control channels (CoreMIDI, Accessibility, AppleScript, CGEvent, MCU, Scripter, MIDI Key Commands — Sources/LogicProMCP/Channels/Channel.swift) behind one MCP interface, then wrapping every high-risk operation in explicit state, confirmation, and verification contracts.

The result is not "screen automation with prompts." It is a structured server for DAW agents: tools mutate, resources read, evidence is labeled, and uncertain outcomes stay uncertain instead of being reported as success.

You: "Make a 4-bar techno loop in A minor at 140 BPM"

MCP client → logic_tracks.record_sequence { bar: 1, tempo: 140, notes: "45,0,95;57,107,95;45,214,95;..." } MCP client → logic_tracks.set_instrument { index: 0, path: "Electronic Drums/Roland TR-909" }

Logic Pro MCP: region imported, instrument routed, readback exposed through resources.

At a Glance

Counts below are read from the current source tree with the command shown; re-run it after pulling to reproduce.

If this project helps you make music with Claude, Cursor, Codex, or any MCP client, star the repo. It helps the project reach more Logic Pro users and maintainers.

Want to contribute? Start with the Contributing Guide and the open issues. Many docs, examples, validation tests, and CLI-message improvements do not require Logic Pro.

Why It Exists

Most Logic Pro automation attempts fall into one of three traps:

1. Prompt-only recipes that drift away from the real tool surface.
2. Keyboard macro automation that can click the wrong target and still look successful.
3. Single-channel control that can write to Logic but cannot reliably read what Logic actually did.

Logic Pro MCP uses a different model. It routes each operation to the strongest available channel, exposes live state through MCP resources, and forces callers to handle three outcomes: confirmed, uncertain, or failed.

What It Controls

Agent-Grade Surfaces

Tools are for actions and local artifact checks. The public write surface is intentionally small: logic_transport, logic_tracks, logic_mixer, logic_plugins, logic_midi, logic_edit, logic_navigate, logic_project, and logic_system. logic_audio is read-only and verifies exported files after Logic writes them.

Resources are for state. Clients should read logic://transport/state, logic://tracks, logic://mixer, logic://project/info, logic://project/audit, logic://project/cleanup-plan, logic://midi/ports, and related resources instead of burning tool calls on polling.

Evidence is separated from claims. The README points to release evidence, current-main verification, and live media artifacts instead of implying that a successful command equals a verified Logic state.

Trust Model

• Honest Contract envelopes: mutating operations return State A confirmed, State B uncertain with a reason, or State C failure with an error.
• Verified plugin apply-back: logic_plugins.* uses HC v2 (hc_schema: 2) and returns State A only after project identity, target track, physical insert slot, plugin identity, and readback all agree.
• Fail-closed targets: dangerous mixer, marker, track, MIDI import, and plugin operations require explicit targets and validation.
• Confirmation levels: destructive/project and plugin insertion flows require explicit confirmation metadata before execution.
• Provenance labels: read surfaces expose source, freshness, and evidence labels instead of forcing clients to guess.
• Installer hardening: Homebrew pins SHA256; the shell installer refuses to run without explicit hash/team pins unless same-origin provenance is explicitly allowed.
• Release honesty: v3.19.0 is the current stable line, and README claims stay tied to shipped artifacts, release-tree tests, or explicitly linked live evidence.

Locales

Logic Pro ships ten UI locales, and this server tracks all ten as the default scope (not just en/ko): de-DE, en-US, es-ES, fr-FR, it-IT, ja-JP, ko-KR, pt-BR, zh-CN, zh-TW. UI-string matching (menu items, header labels, dialog buttons) lives in docs/locale/ui-labels.json, the canon index every locale-sensitive locator derives from.

Each of the 203 canon labels carries a per-locale coverage: measured (read from a live, running Logic window and recorded with the host, date, and AX role/attribute), derived (computed from Apple's own .strings tables by Scripts/derive_label_variants.py, not read from a live window), unmeasured, or retired (one label, headerPanHint, superseded by sliderPanHint and kept in t