Independent MCP evidence
Get independent selection evidence for your MCP.
Get an independent tool-selection evaluation for your MCP. We freeze the benchmark before measurement, preserve misses, and produce a reproducible evidence report.
US$249 introductory price. Payment is requested after qualification. You receive a concise report with evidence-backed findings and recommendations.
The offer
Independent MCP Selection Evaluation
Agent Trust Lab evaluates whether a model selects the intended MCP tool across a frozen benchmark and produces an auditable evidence report. This is a paid evaluation service, not certification.
What you receive
A concrete evidence package
- Independent tool-selection evaluation for one MCP/server
- Tested routing scenarios and a benchmark frozen before measurement
- Observed ambiguity or failure cases, with preserved semantic misses
- Evidence-backed findings with provenance and an immutable measurement receipt
- Recommendations tied to the observed selection behavior
- A concise human-readable final report
A public Agent Trust Lab evidence profile may be considered after review only when you explicitly choose eligibility. Publication is never guaranteed or automatic.
Sample evaluation
See the shape of a useful finding
MCP/server: Example Scheduling MCP (fictional)
Scenario: Find an available meeting slot for two participants.
Observed selection issue: The model selected list_events before find_free_time, although the requested outcome was availability.
Severity / impact: Medium — the first call may return calendar entries without resolving an available slot.
Evidence: A frozen scenario, observed first-tool selection, preserved miss, and version-bound measurement receipt.
Recommendation: Clarify the availability boundary in find_free_time and distinguish it from event listing in the tool descriptions.
Example paid result. This is a fictional format example, not a customer result or guarantee. Before/after improvement is reported only when it is actually measured.
Who this is for
For teams responsible for tool selection
This is for MCP server developers, teams publishing multiple tools, and agent-platform or integration teams seeing ambiguity in model tool selection. It is not general AI consulting or a certification.
How it works
Five steps from request to report
- Submit your endpoint, repository, and project details
- We review qualification and confirm the test scope
- Payment is requested after qualification is accepted
- We freeze the benchmark and run the selection measurement
- We deliver the evidence report
Scope, model, case count, price and delivery date are confirmed in writing after qualification, before payment. A description rewrite and a measured before/after comparison are separately agreed; improvement is never guaranteed.
Scope and limitations
What this evaluation does not certify
It measures tool-selection evidence only. It does not certify MCP security, business correctness, uptime, regulatory compliance, safety, future model performance, or overall product quality. It is not certification, and a declined submission is not evidence that an MCP is poor quality.
Qualification
A useful evaluation needs a defensible test surface
- Reachable public MCP or auditable version-bound source
- Sufficiently stable tool schema
- At least two meaningful tool choices where selection evaluation makes sense
- External or defensible gold construction possible
- No prohibited or destructive evaluation requirement
- No authentication barrier unless specifically arranged
We may decline requests that cannot be evaluated safely or defensibly.
Start with qualification
Request an evaluation
Share only public project information. The required fields are your contact, project identity, and public endpoint or repository. The short description is optional.
After submission: qualification review, payment request if accepted, evaluation, then report delivery. Do not submit API keys, secrets, passwords, or private credentials.