Documentation

Use HTTP, MCP, or OpenAPI. Discovery and beta execution require no account or API key.

Access

Claude and ChatGPT expose the same 37 direct tools across 24 modules through separate free budgets. HTTP and core MCP are also free while PAYMENTS_MODE=disabled. Side effects require a random UUID v4 so identical retries are safe.

View integration and access details →

Reliability Sidecar Contract v1

Use the portable contract to classify an external action, preserve ambiguous outcomes, reconcile before retrying, and report the achieved guarantee without claiming a cross-plugin transaction.

View the machine-readable contract →

Effectiveness and status

Status publishes release, freshness, and current self-test evidence. Effectiveness explains the measurement method and adds privacy-thresholded external reliability aggregates only after the clean baseline is activated.

View effectiveness evidence →

External beta feedback

Invited beta partners can attach enum-only outcomes to opaque execution receipts with a privately issued partner code. The service accepts no task text, prompts, payloads, URLs, credentials, or personal data, retains records for 90 days, and publishes aggregates only after the privacy threshold is met.

Submit bounded beta feedback →

Safe invocation sequence

  1. Search the catalog with a concrete intent.
  2. Inspect the selected tool and its negative cases.
  3. Validate input before any payment or side effect.
  4. Reuse the same idempotency identity only for an identical retry.
  5. For workflow checkpoints, use a fresh identity for each new generation or recovery decision.

Missing capability intake

Submit a bounded request to POST /v1/capability-requests, then poll it with the returned token. Never include credentials, personal data, or conversation history.

View request schema →