⭐⭐⭐ pi-webcam ⭐⭐⭐

pi extension that lets your coding agent see through your webcam.

Capture webcam snapshots from the terminal agent — no browser, no MCP server, no UI. Just /camera take and the image is in context.

What it does

• /camera take — captures a snapshot from your default webcam
• /camera list — lists available camera devices
• Works with text-only models when paired with pi-vision-handoff (or any vision model natively)
• Cross-platform: macOS (AVFoundation), Linux (v4l2), Windows (DirectShow)

Installation

Via pi (recommended)

pi install git:github.com/tkoyama010/pi-webcam

To pin a specific version:

pi install git:github.com/tkoyama010/[email protected]

Manual

git clone https://github.com/tkoyama010/pi-webcam.git ~/.pi/agent/extensions/pi-webcam

Or symlink:

git clone https://github.com/tkoyama010/pi-webcam.git ~/src/pi-webcam
ln -s ~/src/pi-webcam/src/index.ts ~/.pi/agent/extensions/pi-webcam.ts

Usage

> /camera take
# Captures from default camera, saves to /tmp/pi-webcam-*.jpg

> /camera take --device 1
# Uses a specific camera device

> /camera list
# Lists all detected camera devices

After capture, the agent reads the image automatically and can describe it, answer questions, or take action.

Requirements

• ffmpeg — must be installed and in PATH
  • macOS: brew install ffmpeg
  • Linux: apt install ffmpeg
  • Windows: download from ffmpeg.org

How it works

This extension shells out to ffmpeg to capture a single frame from the system webcam. The image file path is returned in the tool result, and the agent's read tool picks it up. On text-only models, pi-vision-handoff (or a similar vision proxy) describes the image so the model can reason about it.

No web server. No UI. No extra ports.

Compared to mcp-webcam

| | pi-webcam | mcp-webcam | |---|---|---| | Type | Native pi extension | MCP server | | UI | None | Browser (React) | | Setup | Install extension | Run server + configure MCP client | | Transport | Direct ffmpeg call | MCP protocol over HTTP/stdio | | Vision | Any model (handoff if needed) | Requires multimodal client | | Footprint | ~50 lines | ~30+ dependencies |

If you already have pi and ffmpeg, this is a 10-second install.

License

MIT