Frigate

Frigate is a network video recorder with object detection. Its Generative AI feature writes object and review descriptions from snapshots, and it has a chat.

Upstream: source, docs.

Plugboard options

Give Frigate its own key and a vision alias, because descriptions are image-based. Docker needs a reachable address (Serving other machines).

{
  "options": {
    "listen": { "host": "0.0.0.0", "port": 4010 },
    "keys": { "frigate": "{env:PLUGBOARD_KEY_FRIGATE}" },
    "models": ["anthropic/claude-sonnet-4-6"],
    "aliases": { "vision": "anthropic/claude-sonnet-4-6" },
    "apps": { "frigate": { "allowedModels": ["vision"] } }
  }
}

Connect

Add a provider under genai in config.yml:

genai:
  plugboard:
    provider: openai
    base_url: http://host.docker.internal:4010/v1
    # Read from the environment: FRIGATE_PLUGBOARD_KEY, the plugboard `frigate` app key.
    api_key: "{FRIGATE_PLUGBOARD_KEY}"
    # An alias from plugboard's `aliases` option. It must accept images.
    model: vision
    roles:
      - chat
      - descriptions

semantic_search:
  enabled: true
  model: jinav1
  model_size: small

plugboard is a name you choose. Frigate expands {FRIGATE_PLUGBOARD_KEY} from the environment, so set it to the plugboard key. Do not give the provider the embeddings role.

The file is in examples/self-hosted/frigate/.

What works

• Descriptions (the descriptions role), from images and text.
• Chat (the chat role).

Limits

• Semantic search stays on Frigate's local Jina model. Plugboard serves no embeddings.
• Plugboard reads data: image URLs. It fetches a plain http(s) image URL only when remoteImages is on, and then only public addresses. Turn it on only if Frigate sends URLs.
• Snapshots go to the provider your OpenCode server is signed in to, so they leave your network.