webcam-mcp

An MCP server that gives a coding agent eyes and ears on the host machine: webcam stills, microphone speech-to-text, and optional text-to-speech replies.

Everything runs locally. ffmpeg captures frames, faster-whisper transcribes on CPU, openWakeWord handles the wake word. No audio, video, transcript, or image is sent anywhere. The only network call the server ever makes is a one-time model download from Hugging Face the first time you transcribe something.

Why it exists: a coding agent that can read your screen but cannot see you or hear you is missing a whole side of a conversation. This closes that gap for any MCP client on a machine with a camera and a microphone.

Tools

Related MCP server: Voice MCP

Setup

Requires Python 3.11+, ffmpeg on PATH, and uv.

uv sync

Run standalone to confirm it starts:

uv run server.py

It sits waiting for MCP stdio input, so it appears to do nothing until a client connects. That is normal. Ctrl+C to exit.

Register it in your MCP client as a stdio command, same pattern as any other server.

Notes earned the hard way

• mcp is pinned below 2. mcp 2.0.0 renamed mcp.server.fastmcp to mcp.server.mcpserver, which breaks the import silently at server start. Stay on the 1.x that actually loads.

• Warmup time is not padding. Webcams open dark and auto-exposure needs a moment. Below about one second the frame comes back black or badly exposed. The default is 1.5 seconds for a reason.

• A failed capture is almost always a device already in use. Zoom, Teams, or a browser tab holding the camera produces an ffmpeg I/O error. The tools detect this and say so. The second most common cause is the OS privacy setting blocking desktop apps from the camera or microphone.

• Whisper model size: base (the default) is the speed and accuracy sweet spot. small is meaningfully better on strong accents at roughly three times the compute. Models cache under the user profile after the first download.

• Models load lazily and stay cached per size for the life of the server process. Reloading per call costs seconds and a lot of RAM.

• The default wake model is openWakeWord's pretrained hey_jarvis. No training, no dataset, it ships in the box. At the default 0.5 threshold it is reliable, and false positives are not a practical concern.

Voice and speech limits

speak uses the legacy SAPI voices by default. They sound robotic. Modern OS natural voices are much better and run locally once installed, and speak picks them up automatically once they exist.

The honest architectural limit is that this is a four-hop pipeline: speech to text, text to model, model to text, text to speech. It is slower than a native audio model, and converting speech to text discards tone, volume, and emotion before the model ever sees the words. A one-hop native audio setup would fix both, but that requires a provider that accepts audio natively. Within a text MCP harness, four hops is the correct design, not a shortcut.

capture_stitched only earns its keep when a second camera is actually connected. Otherwise the second half of the image is a connection spinner.

Output files

Captures default to a temp folder with timestamped names. Pass output_path to send them somewhere durable. record_and_transcribe deletes its WAV afterwards unless you pass keep_audio=True.

Nothing here prunes old captures automatically. Clear that folder occasionally.

Selftest

uv run server.py --selftest

Prints the device list without needing an MCP client, which is the fastest way to confirm the install is sane.

License

MIT. See LICENSE.