frigate-auth-mcp

What

An MCP server with authentication for Frigate NVR: lets an AI assistant (Claude Code, Claude Desktop, Hermes or any other MCP client) look at your cameras, events and reviews - and nothing more.

It logs in to Frigate's authenticated port as a user with the viewer role, so Frigate itself enforces what the assistant may see. Tested with Frigate 0.17.x.

Install

Python 3.11+ and uv. Nothing to install by hand: the MCP client starts the server from PyPI with

uvx frigate-auth-mcp

The latest code from GitHub runs with uvx --from git+https://github.com/adel-s/frigate-auth-mcp frigate-auth-mcp.

Setup

1. In Frigate, create a user for the assistant (Settings -> Users) with the role viewer. If the role lets you limit cameras, give it only the cameras the assistant should see.
2. Add the server to your MCP client. For Claude Code, in .mcp.json:

{
  "mcpServers": {
    "frigate": {
      "command": "uvx",
      "args": ["frigate-auth-mcp"],
      "env": {
        "FRIGATE_URL": "https://frigate.example.com:8971",
        "FRIGATE_USER": "assistant",
        "FRIGATE_PASSWD": "${FRIGATE_PASSWD}"
      }
    }
  }
}

Keep the password out of the file: put it in the environment the client runs in.

Tools

Environment

Safety

• Talks to the authenticated port (8971 by default) only, never to the unauthenticated internal port 5000, which gives full API access to anyone who can reach it.
• Logs in as a Frigate user with the role viewer; the only non-GET call is POST /api/login.
• No tool deletes, exports, edits the config or restarts anything.
• Every tool carries the MCP annotations readOnlyHint: true and destructiveHint: false, so clients can tell that the server only reads.
• The raw Frigate config is not exposed: it carries camera stream URLs with credentials.
• Optional image filter FRIGATE_IMAGE_CAMERAS: only these cameras return images (snapshots, frames). This is the second layer of protection, after the camera limit of the Frigate user's role; by default all cameras are allowed. Use it when the MCP client sends images to a third-party model and some frames, such as indoor cameras, must not leave your home.

Skill

skills/frigate-cameras/SKILL.md tells the assistant how to use the tools: which tool answers which question, text before images, small images, what an empty result or an error means. It is a plain Agent Skill: copy the frigate-cameras folder into the skills folder of your client (for Claude Code .claude/skills/, for Hermes ~/.hermes/skills/<category>/).

Why

The Frigate MCP server I found talks to port 5000 without authentication and exposes tools that delete events, start exports, save the config or restart Frigate. Handing that to an AI model is more trust than a camera system deserves. This server gives the assistant eyes, not hands.

Notes

• Without after, Frigate's /api/review returns only the last 24 hours; list_reviews asks for the last 7 days instead, so "what happened this week" gets a full answer.
• list_events with has_description filters on the GenAI description, which the Frigate API cannot do itself.
• Images default to 480 px and JPEG quality 70 to keep the model's context small; height and quality raise them when a detail is too small to see. Models tend to ask for more than they need: FRIGATE_IMAGE_MAX_HEIGHT and FRIGATE_IMAGE_MAX_QUALITY cap what they get, larger requests are lowered to the limit.
• Errors come back as short plain messages (object not found, login rejected, access denied, Frigate unreachable) without the internal Frigate host name.

Tests

uv run python -m unittest discover -s tests

The tests use unittest and a fake Frigate API on httpx.MockTransport, so they need no running Frigate. They also check that no tool sends anything but GET, apart from the login.

GitHub Actions runs them on Python 3.11 to 3.14 on each push to master and dev/* branches.

Development

If something doesn't work or you need a new feature feel free to open an issue and I'll fix it ASAP!

License

MIT, see LICENSE.