Stripe

1. Discoverability — 9/10

Stripe has one of the best-documented APIs on the internet. Clear resource-based URL structure, detailed parameter descriptions, and consistent naming. An AI agent can reliably infer intent from endpoint names alone.

Gap: No official MCP server yet — agents must infer tool definitions from docs. OpenAPI spec occasionally lags behind new features.

2. Tool Surface — 8/10

Core flows map cleanly to tool definitions. Idempotency keys make agents safe to retry without double-charging. The expand parameter reduces round-trips. Webhooks handle async events well.

Gap: Some flows require multiple sequential API calls (customer → payment method → subscription). An agent-native SDK would bundle these.

3. Auth Simplicity — 9/10

API key auth is the best possible outcome for agent use. No OAuth dance, no session management, no refresh tokens. Restricted keys let you scope exactly what an agent is allowed to do. Test mode keys mean safe development.

4. Response Quality — 9/10

Stripe responses are a model for the industry. Consistent object format, predictable pagination, Unix timestamps. The metadata field on most objects lets agents tag resources with their own context.

Gap: Some error messages are user-facing strings rather than machine-readable codes.

5. Error Handling — 8/10

Structured error objects with type, code, and message. Rate limit headers let agents back off intelligently. Idempotency makes retries safe.

Gap: Some edge cases surface as 500s with vague messages.

What Stripe Should Do

1. Publish an official MCP server — Stripe's API is perfectly shaped for MCP tools. An official server would instantly make Stripe accessible to every Claude, Cursor, and agent framework user.
2. Agent-specific error messages — machine-readable codes, not human-readable strings.
3. Bundle common multi-step flows as single API calls for agent convenience.