Sentinel Desk WebMCP

Sentinel Desk is a standalone, open-source reference implementation for human-approved AI incident response. It demonstrates how WebMCP can let an agent investigate incidents, replay risky behavior, request approval, and activate a runtime guard while keeping the human in control.

This project uses synthetic incidents and a local mock runtime. It does not contain or require Theseus Guard proprietary backend code, production credentials, or customer data.

Live demo

Try Sentinel Desk in a WebMCP-enabled browser:

Open the live demo

What the demo shows

• search_incidents finds incidents by severity, status, tag, or free text.
• inspect_trace exposes evidence and a model trace for a selected incident.
• simulate_guard replays a proposed guard against synthetic historical requests.
• request_guard_approval pauses before any policy-changing action.
• apply_guard only succeeds after approval in the human interface.
• get_audit_report returns the visible provenance timeline.

The app also works in normal browsers as a clickable product demo. In a WebMCP-enabled browser, the tools register through the WebMCP Imperative API (document.modelContext.registerTool).

Local development

npm install
npm run dev

For WebMCP testing, enable chrome://flags/#enable-webmcp-testing in a supported Chrome build, then open the local URL. ChatGPT's in-app browser supports WebMCP without the local flag.

Build

npm run build
npm run preview

Architecture boundary

The demo keeps the policy runtime behind a small adapter boundary. The current implementation is intentionally local and synthetic. A future private integration can implement the same runtime contract without adding proprietary code or data to this repository.

License

MIT. See LICENSE.