Which MCP output does your agent see?

Test whether your agent’s harness passes duplicate tool output to the model. The probe puts independent random markers in text and structured output, then records what the agent reports.

Connect the public MCP

Add this endpoint as a Streamable HTTP MCP server. No sign-in or access token is required.

https://mcp-output-probe.vercel.app/api/mcp

Run and submit

1. Call begin_run and save both run_id and the private run_token.
2. Call probe_both, probe_text_only, and probe_structured_only with both identifiers. Keep using them if the MCP connection reconnects.
3. Call submit_result with both identifiers and all three reports. Use null for values the agent cannot see.

The agent receives a submission receipt. Results stay private until reviewed. result_status accepts the same run ID and token to check publication state.

Give your agent this prompt

Use the MCP output probe. Call begin_run first; supply the harness and model only if explicitly known, otherwise leave them null. Save the returned run_id, run_token, and client-declared identity. Keep run_token private; omit it from the final report and evidence. Call probe_both, probe_text_only, and probe_structured_only once each with that run_id and run_token. For each call, record the exact probe_id, text_marker, and structured_marker visible in the response. Use null for every value not visible; do not infer or invent markers. Call submit_result with the run_id, run_token, and all three reports, using those exact field names. If submission times out, retry with the same run_id, run_token, and unchanged reports. If the MCP connection reconnects, keep using the original run_id and run_token. Report the receipt, publication state, and the markers you saw; say “not visible” for null values. Do not fetch the endpoint or website separately, and do not start another run to recover a missing marker.

Copy the prompt on What the harness? · Connection guide

What the probes return

• probe_both: a different random marker in each output field.
• probe_text_only: text content only.
• probe_structured_only: structured content only.

Recorded probes keep one output per run, so retries return the same markers until submission. Exact marker reproduction demonstrates visibility; a missing report alone does not establish the precise model input. Submission and status receipts never reveal missing markers.

Standalone manual probes are still available by omitting both identifiers. They do not create a submitted result.

Published results · Markdown for agents