Audacity-MCP

---

AudacityMCP connects any MCP-compatible AI assistant to Audacity, giving it full control over audio editing through 144 tools spanning effects, cleanup, mastering, transcription, and more. Talk to your AI assistant and it edits your audio in real-time.

No cloud. No API keys for audio processing. Everything runs locally through Audacity's named pipe interface.

If this is useful to you, a star helps other people find it — that's the whole marketing budget for this project. Want to help keep it maintained? Click the Sponsor badge up top.

Demo

Compatibility: AudacityMCP works with Audacity 3.x. Audacity 4 removed mod-script-pipe — the scripting interface this server drives — and ships no replacement, so AudacityMCP cannot control Audacity 4.

Running Audacity 4? There is a companion project: Audacity4-MCP — 159 tools over a TCP JSON-RPC bridge, covering editing, selection, labels, destructive and realtime effects, and live VST3 parameter control.

It is early alpha, and worth being clear about the setup: because Audacity 4 has no remote-control interface yet, it needs a companion fork of Audacity 4 built from source (Visual Studio, Qt 6.10, CMake, Ninja). That puts it in developer territory for now rather than general use. Audacity's maintainers have confirmed an official command and MCP interface is planned once their internal migration completes; when that lands, the fork should no longer be necessary.

Works With

AudacityMCP works with any AI client that supports the Model Context Protocol:

• Claude Desktop — Anthropic's desktop app
• Claude Code — CLI agent
• Cursor — AI code editor with MCP support
• Any other MCP-compatible client

---

Quick Start

1. Get AudacityMCP

Option A: Click the green Code button above → Download ZIP → extract to a folder

Option B: Clone with git:

git clone https://github.com/xDarkzx/Audacity-MCP.git

2. Run the installer (sets up everything else automatically)

Windows: either double-click install.bat in File Explorer, or — if you're already in a terminal from the git clone step above — just keep going in the same PowerShell/Command Prompt window:

cd Audacity-MCP
.\install.bat

macOS / Linux:

cd Audacity-MCP
bash install.sh

The installer does 3 things: installs audacity-mcp from this local folder (the one you just downloaded/cloned — no PyPI or GitHub fetch involved), enables mod-script-pipe in Audacity, and configures Claude Desktop — no manual JSON editing needed. It asks for confirmation before touching either config file, explains exactly what it's about to do first, and always backs up an existing file before changing it. Want to see everything it would do without changing anything? Add --dry-run: .\install.bat --dry-run / bash install.sh --dry-run.

Note: install.bat/install.sh must be run from inside this folder — they only install the code sitting right next to them, they don't download anything themselves.

If you're not using Claude Desktop, install manually with pip install audacity-mcp-server and add to your client's MCP config:

{
  "mcpServers": {
    "audacity": {
      "command": "audacity-mcp"
    }
  }
}

Check your client's MCP documentation for the config file location.

Client can't find/run audacity-mcp? GUI apps don't always see the same PATH a terminal does. Run python -c "import sysconfig; print(sysconfig.get_path('scripts'))" and use the full path it prints (plus \audacity-mcp.exe on Windows or /audacity-mcp on macOS/Linux) as "command" instead of the bare name.

No install.bat/install.sh, nothing touching your system automatically — three steps, all done by hand:

1. Enable mod-script-pipe in Audacity: Edit → Preferences (Windows/Linux) or Audacity → Preferences (macOS) → Modules → set mod-script-pipe to Enabled → OK → restart Audacity.

2. Install the package: open a terminal (Command Prompt/PowerShell on Windows, Terminal on macOS/Linux) and run pip install audacity-mcp-server — a normal PyPI install, no repo clone needed.

3. Configure Claude Desktop: open Claude Desktop → Settings (gear icon) → Developer tab → Edit Config — this opens claude_desktop_config.json in a text editor. Add this inside the "mcpServers" block, keeping any other servers you already have:

   {
     "mcpServers": {
       "audacity": {
         "command": "audacity-mcp"
       }
     }
   }
   

   Save the file and fully restart Claude Desktop (quit from the system tray, not just close the window).

That's the entire install — see the Installation Guide for the same steps with more detail and per-OS notes.

"audacity" doesn't show up as a tool after restarting? The "command": "audacity-mcp" above only works if that command is on the same PATH Claude Desktop itself uses, which isn't always the case (especially if Claude Desktop was already open when you installed Python). If it doesn't connect, run this in a terminal to find the real install location: python -c "import sysconfig; print(sysconfig.get_path('scripts'))" — then replace "audacity-mcp" above with the full path it prints plus \audacity-mcp.exe (Windows, remember to double every backslash: \\) or /audacity-mcp (macOS/Linux).

3. Start editing

Open Audacity, load some audio, then talk to your AI:

"Clean up this podcast recording"
"Master this track for Spotify, it's EDM"
"Transcribe this and add labels at each sentence"
"Add reverb with a large room, then export as FLAC"

Audacity must be open first. AudacityMCP communicates through Audacity's named pipe — it can't launch Audacity for you.

See the full Installation Guide for detailed setup on all platforms and MCP clients.

---

Why AudacityMCP?

Without AudacityMCP: You manually navigate menus, tweak effect parameters by ear, apply effects one at a time, look up ACX specs, and repeat until it sounds right.

With AudacityMCP: You describe what you want in plain English and the AI handles the rest — picking the right effects, setting industry-standard parameters, and chaining operations together.

	Manual Audacity	With AudacityMCP
Podcast cleanup	5+ steps across different menus, guessing compressor settings	"Clean up this podcast" — one sentence
Music mastering	Research genre-appropriate EQ/compression, apply each manually	"Master this for Spotify, it's hip-hop" — genre-tuned presets
Noise removal	Effect → Noise Reduction → Get Profile → select all → apply	"Remove the background noise" — automatic profiling
Batch operations	Repetitive menu navigation for each operation	Describe the full chain and watch it happen
Transcription	Export audio, use external tool, import results back	"Transcribe this and add labels" — stays in Audacity
Learning curve	Know which effects exist and what parameters to use	Just describe the result you want

AudacityMCP is especially useful for:

• Podcasters who want consistent, professional sound without audio engineering knowledge
• Musicians who need quick mastering with genre-appropriate settings
• Content creators working with interviews, voiceovers, or field recordings
• Anyone who'd rather describe what they want than click through menus

---

What Can It Do?

You:  "Clean up this podcast recording"
AI:   Runs auto_cleanup_podcast → HPF 80Hz → noise reduction → compression → safe loudness check

You:  "Master this track for Spotify, it's EDM"
AI:   Runs auto_master_music style=edm → HPF 30Hz → click removal → compression 2.5:1 → bass +2dB → loudness check

You:  "This is a noisy live recording, fix it up"
AI:   Runs auto_cleanup_live → HPF 100Hz → click removal → noise reduction 18dB → compression 5:1

You:  "Transcribe this interview and add labels"
AI:   Runs transcribe_to_labels → faster-whisper transcription → Audacity labels at each timestamp

You:  "Add reverb to the vocals, then export as FLAC"
AI:   select region → reverb effect → export to FLAC

Features

144 Tools Across 11 Categories

Category	Tools	Highlights
Effects	30	Reverb, echo, pitch shift, tempo change, EQ, phaser, distortion, paulstretch, HPF/LPF, bass & treble, tremolo, wahwah
Cleanup & Mastering	18	Noise reduction, compressor, limiter, 9 one-click pipelines, analysis tool
Editing	13	Cut, copy, paste, split, join, trim, silence, duplicate, undo, redo
Project	12	New, open, save, import/export (WAV, MP3, FLAC, OGG, AIFF)
Track	15	Add mono/stereo, remove, set properties, mix & render, mute/solo, pan, volume
Selection	12	Select all/none/region/tracks, zero crossing, cursor positioning
Transport	7	Play, stop, pause, record, play region, get position
Analysis	6	Contrast, clipping detection, spectrum, beat finder, sound labeling
Generation	5	Tone, noise, chirp, DTMF, rhythm track
Transcription (Experimental)	7	Full/selection transcribe, to labels, to SRT/VTT/TXT, model preload
Labels	19	List/find/edit/delete, delete one labeled take's audio by index, batch add, cut/delete/silence labeled regions, marker export (simple/cue/Podlove), per-segment audio export, import/export

---

Pipelines

AudacityMCP includes 9 one-click pipelines for common audio tasks. Each pipeline is designed to be safe for badly recorded audio — it will never boost your audio dangerously. Pipelines clean up and improve your audio, then you can manually adjust loudness afterward if needed.

How Pipelines Work

1. You tell the AI what you want (e.g. "clean up this podcast")
2. The AI picks the right pipeline and starts it
3. The pipeline runs in the background — you get a job_id back
4. Poll with check_pipeline_status every 15-30 seconds to monitor progress
5. When done, a popup appears in Audacity

Safety rule: Pipelines only reduce peaks if they're too hot. They never boost loudness. If you want to hit a specific LUFS target (e.g. -14 for Spotify), ask the AI to run loudness_normalize after you've checked the results look good.

Pipeline Reference

auto_analyze_audio — Analyze before processing

Measures your audio and recommends the best pipeline. Run this first if you're not sure what to do.

You: "Analyze this audio"
→ Returns: peak level, noise floor, clipping status, recommended pipeline

auto_cleanup_audio — Safe cleanup only

Cleans up noise and artifacts without changing loudness at all. Use when levels are already fine.

You: "Just clean up the noise, don't change the volume"
→ DC offset removal → HPF 80Hz → noise reduction → click removal (optional)

auto_cleanup_podcast — Podcast / voiceover

Professional broadcast processing for spoken word.

You: "Clean up this podcast recording"
→ DC offset → HPF 80Hz → noise reduction 12dB → compression 3:1 → safe loudness check

auto_audiobook_mastering — Audiobook (ACX/Audible)

Targets ACX requirements for audiobook distribution.

You: "Master this for ACX / Audible"
→ DC offset → HPF 80Hz → noise reduction 12dB → compression 2.5:1 → safe loudness check → peak cap -3dB

auto_cleanup_interview — Interview / dialogue

Light touch for conversations — preserves natural dynamics.

You: "Clean up this interview recording"
→ DC offset → HPF 80Hz → noise reduction 8dB → compression 2.5:1 → safe loudness check

auto_cleanup_vocal — Singing / studio vocal

Tuned for vocal recordings with presence EQ for clarity.

You: "Process this vocal recording"
→ DC offset → HPF 100Hz → noise reduction 10dB → compression 3:1 → presence EQ (+3dB treble, -1dB bass) → safe loudness check

auto_cleanup_live — Live / field / noisy recording

Aggressive cleanup for noisy environments. First 0.5s must be ambient noise for profiling.

You: "This is a noisy live recording, clean it up"
→ DC offset → HPF 100Hz → click removal → noise reduction 18dB → compression 5:1 → safe loudness check

auto_master_music — Music mastering

Genre-tuned mastering with 6 presets: edm, hiphop, rock, pop, classical, acoustic.

You: "Master this hip-hop track"
→ HPF 30Hz → click removal → compression 2:1 → bass +3dB treble +1dB → safe loudness check

You: "Master this for a classical album"
→ HPF 30Hz → click removal → compression 1.3:1 (very gentle) → no EQ → safe loudness check

Preset	HPF	Compression	Bass EQ	Treble EQ
EDM	30 Hz	2.5:1 / 80ms	+2 dB	+1 dB
Hip-Hop	30 Hz	2:1 / 100ms	+3 dB	+1 dB
Rock	40 Hz	2:1 / 100ms	0 dB	+1 dB
Pop	35 Hz	2:1 / 80ms	+1 dB	+1.5 dB
Classical	30 Hz	1.3:1 / 200ms	0 dB	0 dB
Acoustic	30 Hz	1.5:1 / 150ms	0 dB	0 dB

auto_lofi_effect — Creative lo-fi / vintage

Apply a warm, vintage lo-fi sound. Presets: light, medium, heavy.

You: "Give this a lo-fi vibe"
→ HPF → LPF (muffled highs) → bass/treble warmth → compression 2:1 → safe loudness check

After a Pipeline: Adjusting Loudness

Pipelines intentionally leave loudness alone (they only reduce if peaks are clipping). To hit a streaming target:

You: "Now normalize this to -14 LUFS for Spotify"
→ AI uses loudness_normalize tool with lufs_level=-14

You: "Normalize to -16 LUFS for podcast"
→ AI uses loudness_normalize tool with lufs_level=-16

Why not do this automatically? LUFS normalization can boost quiet/badly recorded audio by 10-20 dB, which blows it out. By separating cleanup from loudness, you get to check the results before the final loudness step.

---

Local Transcription (Experimental)

This feature is experimental and requires separate setup. Everything else works without it.

Powered by faster-whisper — runs entirely offline, your audio never leaves your machine:

• 5 model sizes: tiny, base, small, medium, large-v3
• Transcribe full audio or just a selection
• Export as SRT, VTT, or plain text subtitles
• Auto-add Audacity labels at each spoken segment
• Language detection or specify 99+ languages

Setup required before first use: See Transcription Setup for installation steps.

Want GPU acceleration (10-20x faster)? Run audacity-mcp-setup-gpu — or download setup_gpu.bat/setup_gpu.sh and double-click/run it if you'd rather not use a terminal. Detects your GPU, installs what's needed, and verifies it actually works. NVIDIA GPUs only (AMD/Intel graphics and macOS aren't supported by the transcription backend at all — any NVIDIA card works, GeForce isn't a requirement, just having an NVIDIA GPU is). No NVIDIA GPU? CPU works fine, just slower on long files.

Script says it worked but transcription is still slow/on CPU? That means Claude Desktop is launching audacity-mcp from a different Python than the one the script just verified — see the fix (a manual config edit, five minutes).

---

What's New — v0.1.3

32 new tools (99 → 131), pipeline tuning fixes, and live-tested against Audacity.

• New effects: Reverse, Invert, Repair, AutoDuck, NotchFilter, VocalReduction, AdjustableFade, StudioFadeOut, CrossfadeClips, CrossfadeTracks, ClipFix, SlidingStretch, Tremolo
• New editing: Split (in place), SplitCut, SplitDelete, Disjoin + renamed old split → edit_split_new
• New tracks: StereoToMono, MixAndRenderToNew, MuteAll, UnmuteAll, Resample, AlignEndToEnd, AddLabelTrack
• New selection: CursorToTrackStart/End, CursorToProjectStart/End, SelectCursorToTrackEnd
• New project: EditMetadata, ImportMIDI
• New labels: RegularIntervalLabels
• Pipeline fixes: ACX peak cap -3.0→-3.5dB, live NR 18→12dB, podcast comp 10ms/1s→30ms/200ms, interview release 1s→200ms
• Bug fix: effect_repair now uses long timeout (Audacity shows popup on invalid selection)
• Validation: Added missing range checks on AutoDuck, VocalReduction, SlidingStretch, Resample

---

Troubleshooting

mod-script-pipe Not Enabled

The installer enables this automatically, but if it didn't work (e.g. Audacity was never opened before), enable it manually:

1. Open Audacity
2. Go to Edit → Preferences (Windows/Linux) or Audacity → Preferences (macOS)
3. Click Modules in the left sidebar
4. Set mod-script-pipe to Enabled
5. Click OK and restart Audacity

Connection Issues

Problem	Fix
"Pipe not found"	Open Audacity first. Make sure mod-script-pipe is enabled (see above). Restart Audacity after enabling.
"Pipe timeout"	Audacity is busy. Wait and retry — the late reply is drained in the background and the next command waits for it (or reports "still busy"). Some effects take minutes on long files.
Connection works once then fails	The pipe disconnected (Audacity crash or restart). Just try again — AudacityMCP auto-reconnects.
"Access denied" (Windows)	Audacity and your AI client must run as the same user. Don't mix admin and non-admin.

Pipeline Issues

Problem	Fix
Pipeline blows out / clips the audio	This shouldn't happen anymore — pipelines only reduce peaks, never boost. If it does, undo (Ctrl+Z) and report the issue.
"A pipeline is already running"	Only one pipeline can run at a time. Use check_pipeline_status with your job_id to monitor the current one.
Pipeline finishes but audio is too quiet	That's by design — pipelines don't boost. Ask the AI: "Normalize to -14 LUFS" after checking results.
Noise reduction sounds metallic/warbled	The first 0.5 seconds of your track must be pure silence/room noise for profiling. If it's not, trim to add silence or use auto_cleanup_audio with remove_noise=False.
Pipeline step failed (in warnings)	Individual steps can fail without stopping the pipeline. Check the warnings field in check_pipeline_status for details.

Audio Quality Tips

Want	Do This
Remove background noise	Make sure the first 0.5s of your track is pure room tone (no speech/music). The pipeline uses this to build a noise profile.
Fix clipping	Run auto_analyze_audio first. If it detects clipping, use auto_cleanup_audio before other pipelines.
Hit -14 LUFS for Spotify	Run a cleanup pipeline first, check the results look good, then ask the AI to apply loudness_normalize at -14 LUFS.
Hit -16 LUFS for podcast	Same approach — cleanup first, LUFS second.
ACX audiobook compliance	Use auto_audiobook_mastering. It targets RMS -20 dB with a -3.5 dB peak cap (safety margin).
Quick cleanup without changing volume	Use auto_cleanup_audio — it only removes noise and artifacts, no loudness changes.

General Issues

Problem	Fix
"No module named faster_whisper"	Run pip install faster-whisper. Transcription is optional — everything else works without it.
Model download fails	Check internet and retry. Models cache locally after first download.
Pipes missing in /tmp (macOS/Linux)	Check Audacity is running and mod-script-pipe is enabled. Check Audacity's console for errors.
Using Snap or Flatpak Audacity on Linux	Both sandbox /tmp, so AudacityMCP auto-detects the real pipe location inside Audacity's mount namespace — no setup needed. If auto-detection fails (containers, restrictive ptrace policies, unusual sandboxing), set the AUDACITY_PIPE_DIR environment variable to the directory holding the pipes and it'll be used directly.
Installer says "Audacity config not found" but Audacity definitely runs	If you're running a portable Audacity (a Portable Settings folder next to the executable), it keeps audacity.cfg there instead of the normal per-OS location, so the installer can't find it. Enable mod-script-pipe manually instead (Preferences → Modules).
install.bat said it configured Claude Desktop, but Audacity never shows up as a tool	Claude Desktop installed via the Microsoft Store redirects its config into an isolated per-package folder — older install.bat versions only wrote to the standard %APPDATA%\Claude\ path, which the Store build never reads

---

Architecture

┌──────────────┐     stdio      ┌──────────────┐   named pipe   ┌──────────────┐
│  MCP Client  │◄──────────────►│ AudacityMCP  │◄──────────────►│   Audacity   │
│(AI assistant)│    (JSON-RPC)  │   FastMCP    │  (commands)    │              │
└──────────────┘                └──────────────┘                └──────────────┘
                                       │
                                       ├── audacity_mcp/main.py          (entry point)
                                       ├── audacity_mcp/audacity_client.py (pipe I/O)
                                       ├── audacity_mcp/tool_registry.py  (auto-loader)
                                       └── audacity_mcp/tools/            (11 modules)

Key Design Decisions

• Named pipes, not TCP — Direct IPC to Audacity's mod-script-pipe. No network exposure, no port conflicts.
• Zero exec/eval — Every operation maps to a static handler with input validation. No arbitrary code execution.
• Cross-platform — Windows uses Win32 API via ctypes, Unix uses standard file I/O.
• Async throughout — All tool handlers are async. Blocking pipe I/O runs in an executor pool with configurable timeouts.
• Safe pipelines — Pipelines measure audio before making loudness decisions. They only reduce, never boost.
• Dynamic tool registration — Drop a module in audacity_mcp/tools/, export a register(mcp) function, and it's automatically discovered.

Project Structure

AudacityMCP/
├── audacity_mcp/
│   ├── main.py                 # FastMCP server entry point
│   ├── audacity_client.py      # Cross-platform named pipe client
│   ├── tool_registry.py        # Auto-discovers and registers tool modules
│   └── tools/
│       ├── analysis_tools.py   # Audio analysis (contrast, spectrum, beats)
│       ├── cleanup_tools.py    # Noise reduction, mastering, 9 pipelines
│       ├── edit_tools.py       # Cut, copy, paste, split, join, trim
│       ├── effects_tools.py    # Reverb, echo, pitch, EQ, filters
│       ├── generate_tools.py   # Tone, noise, chirp, DTMF generation
│       ├── label_tools.py      # Label management
│       ├── project_tools.py    # Project/file operations
│       ├── selection_tools.py  # Selection and cursor control
│       ├── track_tools.py      # Track management
│       ├── transcription_tools.py  # Whisper-based transcription
│       └── transport_tools.py  # Playback and recording control
├── audacity_mcp_shared/
│   ├── constants.py            # Pipe paths, timeouts, allowed formats
│   ├── error_codes.py          # Typed error codes (pipe/command/validation)
│   └── pipe_protocol.py        # Command formatting and response parsing
├── tests/                      # 60 tests
├── docs/
│   ├── INSTALLATION.md         # Detailed setup guide
│   └── TOOLS.md                # Complete tool reference
└── pyproject.toml

---

Development

# Install dev dependencies
pip install -e ".[dev]"

# Run tests
pytest tests/ -x -q

# Run lint (real bugs + security patterns, not style/formatting)
ruff check .

Adding New Tools

1. Create a module in audacity_mcp/tools/ (or add to an existing one)
2. Export a register(mcp: FastMCP) function
3. Define your tools with @mcp.tool() decorators
4. That's it — the tool registry auto-discovers it on startup

# audacity_mcp/tools/my_tools.py
from mcp.server.fastmcp import FastMCP
from audacity_mcp_shared.error_codes import AudacityMCPError, ErrorCode


def register(mcp: FastMCP):
    from audacity_mcp.main import client

    @mcp.tool()
    async def my_custom_effect(intensity: float = 0.5) -> dict:
        """Apply my custom effect to the selected audio."""
        if not 0 <= intensity <= 1:
            raise AudacityMCPError(ErrorCode.VALUE_OUT_OF_RANGE, "intensity must be 0-1")
        return await client.execute_long("MyEffect", Intensity=intensity)

See CONTRIBUTING.md for full guidelines.

---

Community

Turning this into a tool every Audacity user reaches for takes more than one person. If you're using AudacityMCP — even just trying it out — come join the Discord: share what you built, report what's broken, suggest what's missing, or just hang out with other people doing AI-driven audio editing. Communities grow one person telling another this exists, so if you know someone who'd get value out of this, send them the link.

---

Support

If AudacityMCP has saved you time or helped with your audio projects, consider sponsoring:

Your support helps keep this project maintained and free for everyone.

---

Hiring / Consulting

Need a custom MCP server, a native plugin/IPC integration, or a local AI pipeline built for your product? This repo and its sibling projects — Reaper-MCP, SongForge-MCP — are public, working examples of exactly that. Open to contract work: [email protected].

---

Documentation

• Installation Guide — Detailed setup for Windows, macOS, Linux
• Tool Reference — Complete reference for all 144 tools with parameters and ranges
• Transcript-Based Editing — Workflow guide for podcasts, interviews, and lectures (label-driven editing)
• Contributing — How to add tools and contribute
• Changelog — Version history and release notes

Acknowledgments

Thanks to the following for reported bugs, ideas, and feedback that improved this project:

• @Harsha-Jay-S — diagnosed the effect_change_pitch/effect_change_speed bug (#15) with measured proof against real Audacity, pinpointing the exact root cause.

License

Apache License 2.0 — see LICENSE for details.

Built by Daniel Hodgetts • 𝕏 @daehonz1