IP摄像头管理MCP服务

Dahua MCP Server

MCP server for managing Dahua/Amcrest IP cameras via their CGI HTTP API. Supports multiple cameras from a single instance, with both stdio and HTTP transports.

Quick Start

1. Create a cameras config file

Create ~/.config/dahua-mcp/cameras.yaml (or cameras.json) with your camera credentials. This is the default config location — no env var needed.

JSON (cameras.json):

{
  "cameras": [
    {
      "name": "front-door",
      "host": "192.168.1.108",
      "port": 80,
      "username": "admin",
      "password": "secret"
    },
    {
      "name": "backyard",
      "host": "192.168.1.109",
      "port": 80,
      "username": "admin",
      "password": "secret"
    }
  ]
}

YAML (cameras.yaml):

cameras:
  - name: front-door
    host: 192.168.1.108
    port: 80
    username: admin
    password: secret
  - name: backyard
    host: 192.168.1.109
    port: 80
    username: admin
    password: secret

Each camera entry supports:

2. Add the MCP server

Option A: stdio (Claude Code / Claude Desktop)

Using the Claude Code CLI (cameras config auto-discovered from ~/.config/dahua-mcp/):

# From PyPI
claude mcp add dahua-mcp -- uvx dahua-mcp

# From a local clone
claude mcp add dahua-mcp -- uv run --directory /path/to/dahua-mcp dahua-mcp

# With a custom config path
claude mcp add dahua-mcp -e DAHUA_CAMERAS_CONFIG=/path/to/cameras.json -- uvx dahua-mcp

Or manually add to your Claude Code settings (~/.claude.json) or Claude Desktop config:

{
  "mcpServers": {
    "dahua-mcp": {
      "type": "stdio",
      "command": "uvx",
      "args": ["dahua-mcp"]
    }
  }
}

Or run from a local clone:

{
  "mcpServers": {
    "dahua-mcp": {
      "type": "stdio",
      "command": "uv",
      "args": ["run", "--directory", "/path/to/dahua-mcp", "dahua-mcp"]
    }
  }
}

To override the config location, add an env block:

{
  "env": {
    "DAHUA_CAMERAS_CONFIG": "/custom/path/cameras.yaml"
  }
}

Option B: Docker container (HTTP transport)

Run the container with your config file volume-mounted:

docker run -d \
  -v /path/to/cameras.json:/config/cameras.json:ro \
  -p 8000:8000 \
  ghcr.io/brianegge/dahua-mcp:latest

Then add the HTTP server to your MCP client config:

{
  "mcpServers": {
    "dahua-mcp": {
      "type": "http",
      "url": "http://localhost:8000/mcp"
    }
  }
}

To use a YAML config with Docker:

docker run -d \
  -v /path/to/cameras.yaml:/config/cameras.yaml:ro \
  -e DAHUA_CAMERAS_CONFIG=/config/cameras.yaml \
  -p 8000:8000 \
  ghcr.io/brianegge/dahua-mcp:latest

To add bearer token authentication:

docker run -d \
  -v /path/to/cameras.json:/config/cameras.json:ro \
  -e MCP_HTTP_BEARER_TOKEN=your-secret-token \
  -p 8000:8000 \
  ghcr.io/brianegge/dahua-mcp:latest

{
  "mcpServers": {
    "dahua-mcp": {
      "type": "http",
      "url": "http://localhost:8000/mcp",
      "headers": {
        "Authorization": "Bearer your-secret-token"
      }
    }
  }
}

Environment Variables

All settings beyond camera credentials are configured via environment variables:

Available Tools

Discovery

System Info

Config Read

Config Write

System Control

Storage / Recordings

Logs

Firmware quirks

The device protocol lives in aiodahua, which this server is a thin MCP layer over. That is where the digest/basic auth negotiation, the CGI parsing and the firmware quirks are implemented and tested, so the Home Assistant integration and this server share one implementation instead of two.

Behaviours confirmed against real hardware, handled for you:

• Clearing a config field. Sending Key= with an empty value returns OK and changes nothing. Pass "" to set_config and it is sent as a single space, which the device trims back to "".
• & cannot appear in a config value. The firmware percent-decodes the whole query string before splitting it on &, so an encoded & still terminates the value — the request either returns HTTP 400 or the value is silently truncated. set_config raises a clear error instead. #, +, %, =, / and ' are all fine and round-trip correctly.
• used == total on a healthy disk. A recorder pre-allocates the whole drive into fixed-size blocks when it formats, so even a brand new disk reports 100% used. get_storage_info flags this so it is not mistaken for a full disk; use find_recordings to confirm writes are landing.
• find_recordings channels are 1-based. Channel 0 is rejected.
• A missing endpoint is an error body, not a 404. Older firmware answers Bad Request!, 2024 builds answer Not Implemented! with HTTP 501, and some failures arrive as HTTP 200 with an error body that would otherwise parse into a plausible-looking dict.
• Some devices only accept basic auth. Digest is tried first and the fallback is automatic.
• A Lorex E891AB reboots if asked for audio.cgi, dropping HTTP and RTSP for ~105 seconds. The library refuses that request on the brands known to do it.

Development

git clone https://github.com/brianegge/dahua-mcp.git
cd dahua-mcp
uv sync
uv run pre-commit install  # Install ruff format/lint hook
uv run pytest              # Run tests
uv run ruff check .        # Lint

License

MIT License - see LICENSE file for details.