mcp-webcam

In celebration of getting 52 GitHub stars, mcp-webcam 0.2.0 is here! Now supports streamable-http!! No installation required! - try it now at https://webcam.fast-agent.ai/. You can specify your own UserID by adding ?user=<YOUR_USER_ID> after the URL. Note this shared instance is for fun, not security - see below for instructions how to run your own copy locally.

In streamable-http mode multiple clients can connect simultaneously, and you can choose which is used for Sampling.

If we get to 100 stars I'll add another feature 😊.

When run in Streaming mode, if you set an MCP_HOST environment variable the host name is used as a prefix in URL construction, and 5 character UserIDs are automatically generated when the User lands on the webpage.

MCP Server that provides access to your WebCam. Provides capture and screenshot tools to take an image from the Webcam, or take a screenshot. The current image is also available as a Resource.

mcp-webcam supports "sampling"! Press the "Sample" button to send a sampling request to the Client along with your entered message.

Install a recent version of NodeJS for your platform. The NPM package is @llmindset/mcp-webcam.

To start in STDIO mode: npx @llmindset/mcp-webcam. This starts the mcp-webcam UI on port 3333. Point your browser at http://localhost:3333 to get started.

To change the port: npx @llmindset/mcp-webcam 9999. This starts mcp-webcam the UI on port 9999.

For Streaming HTTP mode: npx @llmindset/mcp-webcam --streaming. This will make the UI available at http://localhost:3333 and the MCP Server available at http://localhost:3333/mcp.

You can run mcp-webcam using Docker. By default, it starts in streaming mode:

docker run -p 3333:3333 ghcr.io/evalstate/mcp-webcam:latest

• MCP_TRANSPORT_MODE - Set to stdio for STDIO mode, defaults to streaming
• PORT - The port to run on (default: 3333)
• BIND_HOST - Network interface to bind the server to (default: localhost)
• MCP_HOST - Public-facing URL for user instructions and MCP client connections (default: http://localhost:3333)

# STDIO mode
docker run -p 3333:3333 -e MCP_TRANSPORT_MODE=stdio ghcr.io/evalstate/mcp-webcam:latest

# Custom port
docker run -p 8080:8080 -e PORT=8080 ghcr.io/evalstate/mcp-webcam:latest

# For cloud deployments with custom domain (e.g., Hugging Face Spaces)
docker run -p 3333:3333 -e MCP_HOST=https://evalstate-mcp-webcam.hf.space ghcr.io/evalstate/mcp-webcam:latest

# Complete cloud deployment example
docker run -p 3333:3333 -e MCP_HOST=https://your-domain.com ghcr.io/evalstate/mcp-webcam:latest

If you want a Client that supports sampling try:

Start the mcp-webcam in streaming mode, install uv and connect with:

uvx fast-agent-mcp go --url http://localhost:3333/mcp

fast-agent currently uses Haiku as its default model, so set an ANTHROPIC_API_KEY. If you want to use a different model, you can add --model on the command line. More instructions for installation and configuration are available here: https://fast-agent.ai/models/.

To start the server in STDIO mode, add the following to your fastagent.config.yaml

webcam_local:
   command: "npx"
   args: ["@llmindset/mcp-webcam"]

VSCode versions 1.101.0 and above support MCP Sampling. Simply start mcp-webcam in streaming mode, and add http://localhost:3333/mcp as an MCP Server to get started.

Claude Desktop does NOT support Sampling. To run mcp-webcam from Claude Desktop, add the following to the mcpServers section of your claude_desktop_config.json file:

    "webcam": {
      "command": "npx",
      "args": [
        "-y",
        "@llmindset/mcp-webcam"
      ]
    }

Start Claude Desktop, and connect to http://localhost:3333. You can then ask Claude to get the latest picture from my webcam, or Claude, take a look at what I'm holding or what colour top am i wearing?. You can "freeze" the current image and that will be returned to Claude rather than a live capture.

You can ask for Screenshots - navigate to the browser so that you can guide the capture area when the request comes in. Screenshots are automatically resized to be manageable for Claude (useful if you have a 4K Screen). The button is there to allow testing of your platform specific Screenshot UX - it doesn't do anything other than prepare you for a Claude intiated request. NB this does not not work on Safari as it requires human initiation.

That's it really.

This MCP Server was built to demonstrate exposing a User Interface on an MCP Server, and serving live resources back to Claude Desktop.

This project might prove useful if you want to build a local, interactive MCP Server.

Thanks to https://github.com/tadasant for help with testing and setup.

Please read the article at https://llmindset.co.uk/posts/2025/01/resouce-handling-mcp for more details about handling files and resources in LLM / MCP Chat Applications, and why you might want to do this.