Use case blueprint
Video agents with MCP-style tools and human review boundaries.
A credible video-agent use case starts with read-only evidence tools. MCP-style wrappers can make the interface ergonomic, but the authority boundary must remain explicit.
Implementation blueprint
| Stage | Build | Safety check |
|---|---|---|
| Process | Create video record and materialize artifacts | Cost and detector report visible |
| Expose | Map artifacts to MCP resources | Read-only resources only |
| Answer | Use ask_video with evidence_required=true | No citation, no answer |
| Review | Route risky outputs to human queue | No physical or identity action |
| Audit | Log artifact version, evidence refs, policy result, cost | Trace can be reviewed |
MCP-specific evaluation metrics
Deep pilot note
Do not pilot MCP video tools by giving an agent a camera and asking it to decide. Pilot by selecting one safe video class, one repeated question pattern, one review workflow, and one policy contract. The best first win is a reliable review packet: answer, evidence_refs, timeline spans, confidence flags, blocked actions, and cost context.
The purpose of this page is to make the MCP boundary concrete for builders, buyers, and reviewers. A useful MCP integration is not a promise that an agent can watch everything and decide everything. It is a disciplined interface over already-materialized evidence. Each tool call should be tenant-scoped, read-only by default, evidence-aware, cost-aware, and auditable. If the request asks for identity, intent, physical access, emergency response, employment action, or destructive change, the tool returns a review-required state instead of pretending that a video artifact is a final decision. This keeps the product useful for developers while making it credible for security, legal, and operations teams.