Integrations
RunOnProof is designed to be consumed by software agents and the platforms that run them. An integration can use an API, MCP-compatible tool, marketplace listing, or a platform-specific Tool, but every channel must preserve the same decision meaning.
The invariant contract
No matter which channel you use, preserve:
- exact action and legal company;
- country and coverage;
- named policy and version when applicable;
- evidence references and freshness;
- PROCEED, REVIEW, and STOP semantics;
- conditions and limitations;
- idempotency and replay;
- pricing or payment requirements;
- release and decision identity.
A channel may simplify setup, but it must not simplify away uncertainty.
Choosing an integration style
| Style | Best for | Main responsibility |
|---|---|---|
| Direct API | Product teams controlling their backend | Validate contract, auth, retries, and display |
| Agent Tool | Agents that select capabilities during a workflow | Give the tool narrow instructions and safe state mapping |
| MCP | Reusable agent environments | Preserve typed inputs, outputs, and permission boundaries |
| Marketplace or catalog | Programmatic discovery and purchase | Verify live availability and economic terms before execution |
Recommended integration sequence
- Read the machine-readable catalog.
- Choose one country, one product, and one action.
- Run all synthetic fixtures.
- Map decision states to safe application behavior.
- Test idempotent retries and replay.
- Test payment-required and source-unavailable paths.
- Use controlled real data without economic mutation where supported.
- Enable a narrow production cohort.
- Reconcile decisions, charges, and outcomes.
- Expand only after evidence and operations remain stable.
Tool instructions for an agent
The agent should be told to use RunOnProof before a named company action, supply the strongest available identifier, never treat REVIEW as approval, show limitations, and reuse the same idempotency key for retries of the same intent. It should not ask the tool vague questions such as “Is this company good?”
Relevance AI
The Relevance AI guide provides a complete installation path, Tool configuration, fixtures, daily use, pricing behavior, security, production checklist, maintenance, rollback, and case material.
Error handling
Separate transport errors, contract validation, business decision states, source limitations, and commercial payment requirements. Automatic retry is appropriate only when safe, bounded, and idempotent. Never retry STOP in a loop or turn an exhausted retry into PROCEED.
Data minimization
Send only information necessary for the named action. Avoid placing secrets, credentials, unnecessary personal data, or raw confidential documents in prompts. Use approved storage and references where possible.
What to log
Log the intended action, idempotency key, decision identifier, product and contract version, release identity, result, evidence references, and economic outcome. Do not log authorization secrets or unnecessary sensitive payloads.