Skip to main content
Glama

NUX MG-50Li MCP

Local, open-source MCP server that lets Claude, ChatGPT/Codex and other MCP clients control a NUX MG-50Li through the NUX Audio macOS editor.

The server exposes one scene-aware tone document instead of making an agent perform dozens of blind clicks. A request can contain the preset, S1/S2/S3, tempo, stomp states, signal-chain routing, effect/amp/IR models, every named knob and device settings. The server validates it, orders the changes, applies them, and can leave the result in the live buffer or explicitly save it.

Firmware 1.1.1 preset volume is represented as preset_volume_db (−12 to +12 dB). The Accessibility driver applies it only when the installed NUX Audio build exposes the top-level dB readout and otherwise fails closed with a clear compatibility error.

This project is independent and is not affiliated with or endorsed by NUX or Cherub Technology. NUX and MG-50Li are trademarks of their respective owners.

Implementation status

  • Preset selection and optional save

  • Scene S1, S2 and S3

  • Tempo/BPM, stomp states, and typed A/B/C stomp assignments

  • Typed serial, split A/B + merge, and A-to-OUT/B-to-XLR routing plans

  • NG, CMP, EFX, AMP, Image AMP, power amp, IR, EQ, S/R, DLY, RVB, WAH, MOD and P.L

  • Model-specific knobs plus a forward-compatible arbitrary-control escape hatch

  • One-message import of multilingual tone reports through nux_apply_text

  • Input trim, global EQ, output/USB console, MIDI and other global settings

  • Read-only dry-run plans, UI snapshots and diagnostics

  • A deterministic mock driver for development without hardware

Preset selection, independent S1/S2/S3 chains, tempo, stomp assignments, model/knob changes, all three routing layouts, and the global settings pages have been exercised on a real MG-50Li with NUX Audio 1.1.7. The driver verifies each rebuilt chain slot before continuing and fails closed whenever the requested visible control cannot be found.

The macOS driver uses NUX Audio's Accessibility surface. A transport boundary is kept separate so a verified direct USB-MIDI implementation can be added without changing the MCP tools or tone JSON format. See architecture.

Long tone reports are supported at the MCP-client layer as well as short commands. The report may be in any language or mix prose, headings, and tables. The client semantically reduces the final applicable settings to a ToneSpec; the server then validates and orders that device-only structure. Research notes, alternate guitars, conditional troubleshooting values, and playing advice are not applied as controls. This language-neutral split is intentional: the hardware server never relies on a fixed list of translated words such as “on”, “scene”, or “gain”.

Before any UI mutation, known model names, parameter names, numeric ranges, and enum values are checked against the bundled device catalogue. A bad extraction therefore fails before the selected scene's chain is rebuilt. Unknown future firmware models remain usable and produce an explicit forward-compatibility warning.

Related MCP server: NDSP QC MCP

Requirements

  • macOS 13 or newer

  • Python 3.11+

  • Xcode Command Line Tools (xcode-select --install)

  • NUX Audio installed

  • MG-50Li connected to NUX Audio

Install

With uv:

git clone https://github.com/andwiseb/nux-mg50-mcp.git
cd nux-mg50-mcp
uv sync --extra dev
./scripts/build-bridge.sh
uv run nux-mg50 doctor

Grant Accessibility access to the application that launches the MCP server and, if macOS lists it separately, .build/nux-audio-bridge under:

System Settings → Privacy & Security → Accessibility

You can trigger the macOS prompt with:

uv run nux-mg50 request-access

Open NUX Audio, connect the MG-50Li, then repeat doctor. A ready installation reports "ready": true.

Screen Recording is optional. It is used only to read the painted color of an existing Console mute or pre-existing module bypass before deciding whether a click is needed. Named module parameter switches use absolute drag positions, and applying a complete tone with rebuilt chains does not require Screen Recording. On macOS 14+, grant it to the launcher/helper if you want idempotent readback for existing mute/bypass states.

MCP client configuration

Use the absolute checkout path. Claude Desktop and other stdio MCP clients accept a configuration equivalent to:

{
  "mcpServers": {
    "nux-mg50": {
      "command": "uv",
      "args": [
        "--directory",
        "/ABSOLUTE/PATH/nux-mg50-mcp",
        "run",
        "nux-mg50-mcp"
      ]
    }
  }
}

Restart the client after changing its configuration. The client process itself needs Accessibility permission because it launches the native helper.

Easiest use: paste a tone report

You do not need to write JSON. Paste a tone report into Claude, ChatGPT/Codex, or any capable MCP client and add one sentence describing the destination and persistence:

Apply this report to the currently selected preset and save it. Use the explicit final tables; ignore research notes, alternate guitars and conditional calibration suggestions.

The client calls nux_apply_text with the original document and its semantic ToneSpec extraction. The server validates every known model, enum, control and range before touching NUX Audio, then applies the operations in a safe order. A SHA-256 receipt identifies the exact source text used for the extraction.

Saving is controlled by a separate outer save flag. Text pasted from a report cannot silently grant itself permission to overwrite a preset. If you say “do not save” or do not mention persistence, the result stays in the live buffer. If you name no preset slot, the server never invents one; an explicit save applies to the currently selected slot.

The language understanding comes from the MCP host you are already talking to. The local server needs no OpenAI/Anthropic API key and contains no fragile Turkish/English keyword parser. This also lets it handle long mixed-language prose and tables while the hardware layer remains deterministic and auditable.

Example request

You can say:

Select preset 43A. Make S1 a tight Brit 800 rhythm tone, S2 add more drive and delay, and S3 a louder lead. Set tempo to 112 BPM. Keep the preset unsaved.

The agent should call nux_plan_tone or nux_apply_tone with a structured ToneSpec. A complete example is in examples/complete-tone.json. An example distilled from a long-form tone report is in examples/beggin-17.json. Scene-local changes default to S1 when scene is omitted; pass scene: 2 or 3 explicitly for the other scenes.

CLI dry-run:

uv run nux-mg50 validate examples/complete-tone.json
uv run nux-mg50 apply --dry-run examples/complete-tone.json

MCP tools

Tool

Purpose

nux_doctor

Check NUX Audio and permissions

nux_snapshot

Read visible controls/values for discovery and verification

nux_capabilities

Describe models, parameters, ranges, routing and settings

nux_plan_tone

Validate and order a complete tone without changing anything

nux_apply_tone

Apply the full ToneSpec surface

nux_apply_text

Validate and apply a semantic extraction from pasted prose/tables; records a source receipt

nux_set_scene

Select S1/S2/S3

nux_set_preset

Select a preset and optionally save

nux_set_tempo

Set BPM

nux_set_stomp

Change a stomp-assigned block

nux_assign_stomp

Assign effect blocks to physical switch A, B or C

nux_set_module

Change one module/model/bypass/knob set

nux_set_control

Set a newly added or uncatalogued visible parameter

Safety behavior

  • Changes are not saved by default. Set save: true explicitly.

  • dry_run: true produces the exact ordered operation list without touching the UI.

  • The server never writes guessed bytes to a preset or USB-MIDI endpoint.

  • Chain replacement is destructive only in the selected live scene; preset storage is untouched until save: true is supplied.

  • Unknown firmware controls remain available through nux_snapshot and nux_set_control; the catalogue is guidance, not a version lock.

Development

uv run pytest
NUX_DRIVER=mock uv run nux-mg50-mcp

The native helper is intentionally small and auditable. Build it after editing:

./scripts/build-bridge.sh

Hardware/UI verification instructions are in CONTRIBUTING.md. The current device evidence and remaining checks are tracked in docs/HARDWARE-VERIFICATION.md.

License

MIT

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables MCP clients to programmatically control an Ableton Live 12 session, including tracks, clips, MIDI notes, devices, parameters, mixer, scenes, warping, and rendering, over streamable HTTP.
      1
      MIT
    • F
      license
      Not graded
      quality
      B
      maintenance
      Enables MCP clients to control Neural DSP Quad Cortex devices, including preset recall, block parameter editing, and device diagnostics through conversational commands.
      -
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables local inspection of Apple MainStage state and patch/set information, plus MIDI control such as program changes, CC, and metronome toggle via MCP.
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Lets AI assistants control a NUX MG-30 guitar processor over USB MIDI by picking amp, cab, drive, modulation, delay and reverb models, toggling blocks, and setting knobs by name. It can also read, save, rename, back up and restore presets on the device, and apply whole tone recipes from JSON patch files.
      57 PyPI
      1
      MIT