Explicit tool approval

Check one fixed MCP tool request.

Connection success and tool-output evidence answer different questions. Approve the exact request whose result you want to inspect.

Preview before enabling

  1. Open Configure tool execution, choose a Discovered tool and enter Fixed JSON arguments, optional Output assertions, an interval and a timeout.
  2. Choose Save and preview request. Saving disables execution. Review the saved endpoint, tool and arguments.
  3. Acknowledge side effects and provider costs for this exact request, then choose Enable tool execution. An annotation such as readOnlyHint does not establish safety.
  4. Inspect execution request and output evidence after an eligible call. Inventory results stay separate.

Illustrative example

Suppose the reserved endpoint https://mcp.example/mcp lists a tool named service_info. A fixed request might use:

{"service":"demo"}

For a JSON result whose expected top-level status is ok, an output assertion could be:

{"required_fields":["status"],"equals":{"status":"ok"}}

A completed call and a matching assertion are distinct evidence. A failed call or mismatching field needs review. These names and values are illustrative, not an installed integration or measured result.

Expectations and model advice need their own review

Saved behavioral expectations and reference fixtures are separate from the execution request's top-level JSON output assertions. Deterministic comparison is distinct from model advice. Separately approve local Laya review if you want that advisory signal; external Cloudflare review needs another explicit approval. Local consent does not authorize external transmission.

Model review uses bounded selected evidence; fixed invocation arguments and unnecessary reference fixtures are excluded. An absent or deferred judgment is not a successful assessment. Neither model guarantees correctness. Execution latency covers sending tools/call through receiving and parsing its result, excluding initialization; it does not measure first-token or internal model time.

Worked monitoring examples

How MCP connection monitoring works → · Understand tool checks →

Start Free Current plans