Sentinel Desk

Theseus Guard - Sentinel Desk WebMCP

Project Story

AI systems can fail in ways that are difficult to investigate and even harder to remediate safely. When an incident occurs, security teams need to understand the evidence, replay the risky behavior, evaluate a possible fix, and preserve a clear audit trail. We wanted to explore how agents could help with this workflow without giving them unrestricted authority to change runtime policies.

That idea became Sentinel Desk: an open-source, standalone AI incident response desk powered by WebMCP.

Sentinel Desk lets an agent search incidents, inspect model traces, identify policy violations, simulate a proposed runtime guard, and request human approval. The final policy-changing action remains explicitly gated by the human operator. After approval, the guard is activated and the action appears in the provenance timeline.

The WebMCP layer exposes six tools:

search_incidents
inspect_trace
simulate_guard
request_guard_approval
apply_guard
get_audit_report

These tools are registered through the WebMCP Imperative API using document.modelContext.registerTool. The application also works as a normal web interface, so people can manually explore the workflow even when WebMCP is not available.

The most important thing we learned is that WebMCP is not simply a replacement for clicking buttons. A well-designed tool gives an agent a precise capability, a clear input schema, and an explicit trust boundary. This is especially important for security workflows: reading evidence and proposing a remediation are different actions from activating a policy change.

One of our main challenges was working with an emerging browser API while preserving a reliable human experience. We had to handle WebMCP availability, tool registration, cancellation and permission boundaries, while keeping the application usable in a regular browser. We also needed to make a synthetic runtime feel realistic without implying that it was connected to production data.

Sentinel Desk demonstrates a possible future for AI governance: agents can help investigate complex incidents and prepare remediation, while people retain authority over consequential runtime changes.