Choose a release
Run the known failure first, then prove the fix against the same reviewed boundary.
Behavioral release evidence for WebMCP
Generated tools need proof.
A tool can claim “read only” and still charge a card. Arena compares the protection outcome of the human and agent routes, watches delayed backend effects, and signs a release verdict.
Live boundary audit
Can this WebMCP tool ship?
The agent prepares the release. You approve the exact intent. Arena then exposes the reviewed candidate as a real registered WebMCP tool and will not issue a conclusive proof until that callback runs.
Review the exact intent
Start an audit to prepare the exact release and intent.
No approval WebMCP tool exists. The interface capability is single-use and session-bound.Invoke and observe
No route has executed.
preparingawaiting approvalawaiting webmcp invocationwaiting for effectscompleted
Unverified evidence
Protection boundary, not task completion
Run either preset to compare authoritative human and agent effects.
Generation is the start
Arena executes owned routes and compares recorded outcomes instead of trusting generated descriptions or read-only annotations.
Approval binds exact intent
The capability commits to the release, claimed agent, tool, arguments, target, contract, session, nonce, and expiry.
Return is not settlement
Delayed writes are observed until a terminal watermark and final state are recorded.