AOPX

Where it sits in the directory's order

What can be seen from outside, weighed as the directory publishes.

Where it answers

• https://api.aopx.fr/mcp/streamable-http

The tools it lists (4, as of 2 days ago)

The tool-description rules found nothing in these descriptions. They look for instructions aimed at a model and for hidden characters; they cannot see what a tool does when it runs.

• recommend_provider

  Use when a software agent needs an AOPX recommendation for the SEARCH category instead of hard-coding a single supported provider. Supported modes are ECONOMY and CRITICAL. Returns a primary provider and, when applicable, a fallback provider plus a recommendation_id. AOPX does not execute the search and does not handle payment; the caller executes the provider path externally. After a real execution, call report_outcome with the same recommendation_id. max_cost_eur and max_latency_ms are accepted and recorded by the current v0 interface but are not enforced as hard routing constraints.

• report_outcome

  Use only after a real external execution of a provider path returned by recommend_provider. Records the caller-reported production outcome for that recommendation_id. The recommendation must exist, only one outcome is accepted per recommendation, and provider attempts must match the recommended primary/fallback path. Reported outcomes are production evidence supplied by the caller; they are not described as independently verified outcomes.

• public_stats

  Read aggregate public statistics from the AOPX production pilot, including consultation and reported-outcome activity. Use this for public telemetry, not for provider recommendation. LAB benchmarks, tests, shadow runs and replay events are excluded from these production statistics.

• public_outcomes

  Read recent caller-reported production outcomes from the AOPX public evidence surface. Use this to inspect what was reported after real recommendations were executed, not to request a new recommendation. Provider identities are exposed only through public aliases. LAB benchmarks, tests, shadow runs and replays are not returned here, and reported outcomes are not described as independently verified.