AI security agent
The optional AI security agent is off by default. Local recording, detection, and appearance search do not require it.
An administrator configures Settings → AI agent:
- Choose Anthropic, OpenAI, or an OpenAI-compatible endpoint and the model name.
- Supply the provider's API key. A compatible endpoint is any server that speaks the OpenAI chat API, including a local one, and may not require one.
- Set the daily budget and question ceiling.
- Select the cameras whose pictures may be sent, enable the agent, and save.
- Use Test the connection and review the result. The test is one small paid request to the provider.
The key is encrypted in the credential vault and never displayed again. Changing provider or compatible endpoint removes the old key so it is not sent to the new destination.
The AI agent is a licensed feature. Where the recorder's license does not include it, enabling the agent or the morning briefing, Test the connection, questions, and commissioning proposals are refused. An agent already enabled keeps reviewing events in the background.
What is shared
Each question sends one cropped event picture and a few hundred words of recorder context. It never sends a video stream. Only pictures from selected cameras may be sent. Identity adjudication can include tagged reference sightings alongside the current visit.
By default the recorder context is limited to the selected cameras as well: their names and zones, the rules that watch them, their recorded visits, and the known people, vehicles, and pets they have seen. Cameras that are not selected are not named to the provider, and account names are never sent. A person asking a question learns only about cameras they may already see, so a question from an operator scoped to one camera is scoped the same way. Camera names, zone names, rule names, health messages, and the question itself are marked in the prompt as data copied from the recorder, so text typed into a name is not read by the model as an instruction.
Share records from every camera is a separate switch, off by default. With it on, a question may carry records from every camera the person asking may see, not only the selected ones: camera, zone, and rule names, event times and labels, the names of known people, vehicles, and pets, and account names. Pictures still come only from the selected cameras. Turning it on or off is written to the audit log.
Choose the provider and camera permissions with the site's owner.
Models and refusals
The Anthropic model list offers Claude Opus 5.5 (the default), Sonnet 5.5, Haiku 4.5, and Fable 5.1; any model name the provider accepts may be typed in, and the earlier ids stay priced. When the chosen Anthropic model declines a request on safety grounds, the request is retried on Anthropic's recommended substitute in the same call and the answer is recorded against the model that produced it. A request every model declines is reported as declined and the alert goes out untriaged.
An OpenAI-compatible endpoint may not point at the recorder itself, a link-local address, or any onboarded camera, whether by address or by a name that resolves to one. The name is resolved when the endpoint is saved and again before each call.
What it can do
- Answer questions in Investigation using the recorder's records.
- Help interpret an event and describe the concern.
- Propose zones and rules for individual review in Detection zones.
- Review uncertain appearance matches when Let the agent settle unsure matches is on. The agent sees the visit beside up to three tagged sightings and judges by clothing, build, shape, and color, never by face. Its answer counts toward each camera's recognition floors, but the event's verdict remains unsure until a person rules. It asks a few every five minutes, on the selected cameras, inside the budget.
- Gate a detection rule using Ask the AI agent first, which holds every action in the rule's list until the agent has looked.
Suggestions require review. A proposed zone or rule is not applied simply because the model suggested it. If rule review fails or times out, the rule's actions run as they would have without asking; see action rules.
Budgets
Spending is counted over the last 24 hours, and questions stop when it reaches the daily budget. Questions still waiting on the provider are counted at an estimate until their cost is known, so several questions asked at once cannot each pass the check. Zero means no cap. When a model's price is unknown, the dollar budget cannot bound spending: enter its prices, both above zero, or use the daily question ceiling, which applies only to such models. Cache reads are priced per model where the provider lists a rate. Check the displayed usage and provider billing when commissioning the feature.