WebMCP review surface

The agent investigates. The human authorizes.SAMPLE DATA

The agent may collect evidence, explain documented contract-cost growth, and stage a review. It never recommends a bidder, selects a winner, or awards a contract — and that boundary is the capability set itself, not a warning. Publishing exists only after a human grants one memo: a one-use, self-expiring tool bound to the review on screen.

Detecting the WebMCP API…

How the capability boundary works

Step 1 of 8

The same eight steps run whether a real WebMCP agent drives this page or you click the controls below. Watch which tools exist at each moment: the boundary is the capability set, not a warning the model is asked to respect.

Auto demo
Caller The agent connecting
inspect_review_boundaryon
research_contractor_historyon
get_cost_growth_evidenceon
stage_preaward_reviewon
publish_authorized_memohuman-locked
recommend_bidderabsent
award_contractabsent
Authority The human holds the decision
What happens

Four read-only tools register on an origin-keyed cluster. Two tools a procurement agent might expect, recommend_bidder and award_contract, are never defined anywhere in the codebase.

Load-bearing guard

Origin-Agent-Cluster: ?1
Permissions-Policy: tools=(self)

Safety posture

—

Why this is safe to hand to an agent. The header protections are verified live in your browser; the rest are enforced server-side and in the tool schemas themselves.

Human mandate

what the human authorizes
Subject—
Work type—
Bidders & bids
Allowed actions
    Forbidden actions
      Authorization—
      Statusnone
      Expires—
      Evidence tools (read only)
      “Watch the server model” asks the server's configured model to investigate the staged letting and shows each tool it calls in the receipts — your local model when MODEL=local, the deterministic mock on the public demo. It never publishes.

      Live capability map

      source: —

      Driven by the WebMCP tool registry itself — getTools() after every toolchange — not a shadow list this page maintains.

      Receipts

      newest first

      Registrations and failures, evidence reads, insufficient-history refusals, staging, authorization, publication, and capability removal.