mcp-listen

Give your AI agents the ability to listen

Microphone capture and speech-to-text tools for MCP-compatible agents.

claude mcp add mcp-listen npx mcp-listen

Claude Desktop / ChatGPT Desktop / Cursor / Windsurf / VS Code

{ "mcpServers": { "mcp-listen": { "command": "npx", "args": ["-y", "mcp-listen"] } } }

Compatible with Claude Desktop, ChatGPT Desktop, Cursor, GitHub Copilot, Windsurf, VS Code, Gemini, Zed, and any MCP-compatible client.

- Windows x64
- macOS Apple silicon (arm64)
- Linux x64 and arm64 (glibc)

Intel Mac (darwin-x64) is not supported: Apple has discontinued the platform and no decibri binary is published for it.

Forlist_audio_devicesandcapture_audio:

- Ollamarunning locally
- Whisper GGML model (seeWhisper Model Setup)

[ { "index": 0, "name": "Microphone", "id": "wasapi:{0.0.1.00000000}.{6b187949-26ea-470b-907d-66bf87261530}", "maxInputChannels": 2, "defaultSampleRate": 48000, "isDefault": true }, { "index": 1, "name": "Microphone Array", "id": "wasapi:{0.0.1.00000000}.{b7a6e3e2-a62b-4e92-9320-947c4be98552}", "maxInputChannels": 2, "defaultSampleRate": 48000, "isDefault": false } ]

Theidis stable across reboots and device changes. Theindexis positional and can shift when devices are added or removed, and names are not unique. Preferidwhen selecting a device. In the rare case the host cannot produce a stable id for a device, itsidis an empty string and it can only be selected byindex.

Records audio from the microphone and saves as a WAV file. Records for exactlyduration_msby default, or until the speaker stops talking withstop_on_silence: true.

{ "path": "/tmp/mcp-listen-1712345678901.wav", "duration_ms": 5000, "sample_rate": 16000, "channels": 1, "size_bytes": 160044 }

Example response (stop_on_silence: true):

{ "path": "/tmp/mcp-listen-1712345678901.wav", "duration_ms": 2600, "sample_rate": 16000, "channels": 1, "size_bytes": 83244, "stopped_by": "silence", "speech_detected": true }

Withstop_on_silence,duration_msin the response is the actual captured length, andstopped_bysays how the recording ended:"silence"(the speaker finished),"ceiling"(theduration_msmaximum was reached), or"no_speech_timeout"(nobody spoke for 10 seconds; the WAV is still returned, withspeech_detected: false, so silence is a reported outcome rather than an error). Detection runs per ~100ms audio buffer, so the effective hangover rounds up to the next buffer, and the recording keeps everything from the start of the call through the stop decision: nothing is gated or clipped at speech boundaries, and the audio itself is byte-identical to a fixed-duration capture of the same sounds.

Full voice pipeline: capture audio, transcribe with whisper.cpp, send to Ollama, return the response. Entirely offline. Recording stops automatically when the speaker stops talking; passstop_on_silence: falsefor a fixed-length recording.

{ "transcription": "What is the default port for PostgreSQL?", "response": "PostgreSQL runs on port 5432 by default.", "model": "llama3.2" }

Result outcomes.voice_queryreports five distinct outcomes. Thestructured fields are the contract(isError,speech_detected,transcription); anymessageis a human-readable hint whose wording is not part of the contract, so a caller branches on the fields, never on the prose. The rule is simple:if the pipeline ran, the result is a success (even when it found no words); if a dependency broke, the result is an error.

A caller distinguishes "the user was silent" from "the user spoke but produced no words" byspeech_detected(falsevstrue), both carryingtranscription: null. Non-speech audio never reaches the language model: whisper's non-speech markers ([BLANK_AUDIO],[MUSIC],(silence), and similar) are treated as no usable words rather than sent on as a query.

{ "speech_detected": false, "stopped_by": "no_speech_timeout", "transcription": null, "response": null, "message": "No speech was detected. Ask the user to repeat, or check that the correct microphone is selected." }

Speech, but no transcribable words(speech_detected: true,transcription: null):

{ "speech_detected": true, "stopped_by": "silence", "transcription": null, "response": null, "message": "Speech was detected but could not be transcribed. It may have been too quiet, too brief, or unclear. Ask the user to repeat, a little louder and closer to the microphone." }

Transcription and dependency failures returnisError: truewith the real cause (a missing model, a whisper load failure, Ollama not running, a timeout, or an empty model response), so a caller debugging can tell whether the failure was in capture, transcription, or the language model.

mcp-listen usesdecibrifor cross-platform microphone capture. No ffmpeg, no SoX, no system audio tools required. Pre-built native binaries with zero setup.

mkdir -p ~/.mcp-listen/models curl -L -o ~/.mcp-listen/models/ggml-base.en.bin https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin

mkdir "$env:USERPROFILE\.mcp-listen\models" -Force Invoke-WebRequest -Uri "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin" -OutFile "$env:USERPROFILE\.mcp-listen\models\ggml-base.en.bin"

The model is ~150MB and downloads once. You can also set theWHISPER_MODEL_PATHenvironment variable to a custom directory.
- Install Ollama fromhttps://ollama.com
- Pull a model:ollama pull llama3.2
- Ensure Ollama is running:ollama serve
- A loud transient can register as speech.Silence-stopping decides "speech has started" from the VAD score, so a door slam or a cough can start the countdown and end the recording aftersilence_msof quiet, yielding a short capture of mostly silence. The outcome is visible, not silent: the result reports the actual duration, andvoice_queryreports an empty transcription rather than inventing one. A minimum-speech-duration guard is a candidate refinement.
- voice_queryrequires Ollama running.If Ollama isn't running, the tool returns a clear error message.
- Whisper model must be downloaded before first use.voice_querydoes not download the model itself; the first call requires a pre-downloaded model (~150MB). SeeWhisper Model Setup.
- No streaming.MCP's request/response pattern means the entire recording is captured, then transcribed, then sent to the LLM. No real-time partial results.
- Temp files.capture_audiowrites WAV files to the system temp directory and returns the path, so the file has to outlive the call for the caller to read it. Recordings older than 24 hours are removed the next time the server starts; recordings made since the last restart persist until then.voice_querydeletes its recording as soon as the query completes.

Whisper: "model not found"The whisper model file must be downloaded before first use. SeeWhisper Model Setupfor instructions.

A bridge server connecting Agent Communication Protocol (ACP) agents with Model Context Protocol (MCP) clients.