I want an endpoint audit to tell me what it observed and what it could not verify. A green card is useful only when the evidence supports it. A red warning is useful only when a client actually hits a defect.
The Agent Endpoint Inspector and the agent endpoint rows in the Mega Analyzer now share one check set. The inspector also shows independent dimensions and coverage. The Mega Analyzer keeps its existing grade formula.
Authentication is a boundary
An HTTP 401 means the anonymous request needs authentication. It does not prove that an authenticated client can use the protocol, and it does not prove the server is broken. A 403, timeout or rate limit leaves the result inconclusive. I keep those unknowns visible instead of treating them as passes or missing support.
An HTML page at a guessed /mcp path is not a successful protocol exchange. It may be a marketing page or a fallback route. The inspector calls that claimed-only evidence. If you operate a server, publish the exact endpoint and the authentication requirements in your own documentation.
A card and a service are different evidence
A valid Agent Card can point to an unreachable service. I check its metadata separately from the GET and OPTIONS responses. Neither response proves that an A2A task can finish.
Historical and current cards use different fields. The validator selects the declared revision and checks required fields and types for that revision. An unknown revision stays unknown. These bounded checks are not a full schema certification, and the report says so.
Tool names are claims
A tool named search can describe holding inventory. The inspector labels names and descriptions as read, hold, book, cancel, pay or unknown, with side effects taking precedence over a search name. These are untrusted metadata labels. They never authorize execution.
Public metadata mentioning guest records or personal details creates a private verification task for the owner. It is not evidence that anyone can retrieve those records. Test authorization in your own controlled environment. Do not publish real guest data in a demonstration or relax authentication to make a scanner happy.
The scan never calls tools, reads resources, retrieves prompts or submits A2A tasks. Instructions embedded in descriptions and cards remain data.
Coverage explains the limits
A dimension score counts the checks with a verified pass or failure. Unknown checks reduce coverage. A dimension with no evaluated checks has no score. That is more honest than calling incomplete evidence perfect.
The endpoint scan does not crawl booking pages or retrieve supplier quotes. It cannot verify taxes, cancellation conditions, inventory or availability. Weekly extended stays, tenancy applications and inquiries must not be described as instant hotel reservations without evidence. No MCP or A2A endpoint is required for an informational site.
The evidence JSON download records observation time, target and final URL, method, status, bounded response hash, redacted excerpt, state and limitations. A hash describes captured content, not its truth. Compare scans only when their scope and validator revisions match. Keep downloaded reports private if your own metadata contains sensitive details.
For a concrete repair path, use the endpoint fix guide. Fix confirmed defects first, then verify unknowns through an authorized client. My book The $97 Launch follows the same habit: make the smallest useful improvement you can test.
Fact-check notes and sources
- HTTP semantics, RFC 9110: authentication challenges and HTTP response semantics.
- A2A current specification: Agent Cards, supported interfaces and protocol boundaries.
- A2A 0.3 specification: historical card fields.
- MCP tool annotations: annotations are hints and must not be trusted as authorization.
