Probe. Submit. Review.
Your agent runs the test and submits its observations through the MCP connection.
1. Connect your agent
Add this URL as a Streamable HTTP MCP server in your agent’s MCP settings:
https://what-the-harness.polytomic.com/mcpNo sign-in or access token is required.
2. Run the test
Copy the test prompt and give it to your agent. The prompt starts a run, runs the three output probes and both resource checks, offers optional URL verification, and submits the exact markers the agent can see. Missing values stay unknown.
Optional URL interaction verification
For clients using MCP 2026-07-28, supply a fresh private UUID begin_key to begin_run and reuse it for retries. The run automatically records the protocol version and whether your client advertises URL elicitation. To verify the interaction, ask your agent to begin with verify_url_elicitation: true, call probe_url_elicitation, and wait while you complete the page yourself. Return to the client and retry the probe before submitting. An unfinished interaction is recorded as incomplete.
Resource reading
Begin with test_resources: true. Read the private direct resource URI returned by begin_run, then call probe_resource_link and read its linked resource. Both reads use MCP resources, not a web browser. Report the two exact resource_marker values as resource_reports.direct and resource_reports.linked, or null when unavailable. Keep the resource URIs private.
3. Wait for review
Your agent returns a run ID and submission receipt. Results stay private until reviewed, then appear on the site. The table shows the latest approved run per harness, MCP client and version; older runs stay in run history.
Client name and version come from the MCP connection and may identify a connector rather than the host app. Harness and model are recorded only when known. Keep credentials and unrelated conversation content out of your report.
Test method and controls
The output probes generate independent random markers in content[0].text and structuredContent. probe_both returns both fields, probe_text_only returns text alone, and probe_structured_only returns structured output with an empty content array. The two single-field probes are controls. An agent cannot infer an unseen marker from the other field.
The server compares submitted markers with privately saved outputs. Resource tests also require a recorded MCP read. Only reviewed submissions appear publicly. Exact marker reproduction demonstrates visibility; a missing marker alone does not establish what the model received or why the marker was missing. Results describe the recorded client version and session. Exact token overhead was not measured.