API Design Prover MCP Connector for Claude
A+An AI agent designed an API with GET /users/create. That single endpoint broke HTTP caching for 200 consumer services. No versioning. No error contract. Raw arrays on one route, wrapped objects on another. This tool forces semantic HTTP verbs, explicit versioning strategy, unified response envelopes, bounded pagination, and RFC 7807 error structures — before any code is written.
LLMs design APIs the way junior developers do — they pick verbs randomly, skip versioning, return different shapes on every endpoint, and dump error stack traces to the client. The result: 200 consumer services break when you refactor a single route. Clients write custom parsers for each endpoint. Nobody can paginate. Nobody can cache.
The Problem
AI-generated API designs commit five protocol violations that compound into integration nightmares:
- Verb Abuse — GET /users/create. POST for reads. PUT for partial updates. The agent treats HTTP methods as cosmetic labels. Every misused verb breaks caching, idempotency guarantees, and proxy behavior for every consumer downstream.
- Version Blindness — No
/v1/prefix. No header strategy. No deprecation policy. The first schema change becomes a breaking change for every integrated service. Rollback is impossible. - Shape Chaos — Raw arrays on
/users, wrapped objects on/users/:id, nested envelopes on/users/:id/orders. Three endpoints, three different response contracts. Client SDKs become unmaintainable. - Unbounded Lists —
GET /ordersreturns 47,000 rows. No cursor. No limit. No offset. The mobile client crashes. The CDN times out. The database locks. - Error Void — Status 500 with
{ message: 'Something went wrong' }. No error code. No RFC 7807. No machine-readable structure. The frontend shows a generic toast. Support tickets multiply.
How It Works
API Design Prover validates every contract through 5 Decision Pivots:
- verbsSemantic — GET is read-only and idempotent. POST creates. PUT replaces entirely. PATCH updates partially. DELETE removes. No exceptions. No 'it depends.'
- versioningDefined — URL path (
/v1/), header, or query param — with an explicit deprecation timeline. 'We'll version later' fails this pivot. - responseConsistent — Every endpoint returns
{ data: T, meta?: {}, errors?: [] }. No raw arrays. No naked objects. One shape, every route. - paginationHandled — Collection endpoints use cursor or limit/offset with a hard maximum (e.g., max 100).
GET /userswithout bounds fails immediately. - errorContracted — Errors follow RFC 7807 Problem Details:
{ type, title, status, detail, instance }. String messages are rejected.
The Verdict Matrix
| First Failing Pivot | Verdict | Meaning |
|---|---|---|
| verbsSemantic = false | VERB_ABUSE | HTTP semantics violated. Caching and idempotency broken. |
| versioningDefined = false | VERSION_MISSING | No versioning strategy. First change breaks all consumers. |
| responseConsistent = false | SHAPE_INCONSISTENT | Different response shapes per endpoint. Client SDKs unmaintainable. |
| paginationHandled = false | UNBOUNDED_RESPONSE | Lists return unlimited rows. Memory leaks and timeouts guaranteed. |
| errorContracted = false | ERROR_UNDEFINED | No standard error structure. Consumers cannot handle failures programmatically. |
| All pivots pass | API_PROVEN | Semantic verbs, versioned, consistent shape, paginated, RFC 7807 errors. |
Why It Works
- Tool calls are obligations. The agent cannot claim 'versioning is handled' without naming the strategy. Filling the fields IS the design work.
- Consistency engine catches contradictions. Claiming
responseConsistent=truewhile returning raw arrays on one route triggers rejection with a specific coaching message. - Semantic traps detect vague answers. 'We follow REST best practices,' 'errors are handled by the framework,' or 'pagination will be added later' all trigger automatic rejection.
Related Connectors
Legal Counsel Prover MCP
AI agents cite fabricated statutes, ignore deadlines, and deliver one-sided legal memos. This tool forces rigorous reasoning: identify jurisdiction, cite verifiable law, map procedure, address the opposing argument, connect to the client's facts.
Accessibility Prover MCP
Accessibility Prover is a pre-build validation system that forces WCAG 2.2 AA compliance before code is written. It audits semantic HTML, keyboard navigation, contrast ratios (4.5:1 minimum), screen reader compatibility, and motion safety across 5 structured Decision Pivots. Unlike runtime scanners, it catches div soup, missing focus traps, and inaccessible animations during the design phase — before they reach production and before they violate the European Accessibility Act (EAA 2025).
Context Integrity Prover MCP
AI forgets the original goal and hallucinates scope. This engine is a 6-pivot trap that forces the LLM to prove it hasn't drifted from the user's explicit constraints.
First Principles Prover MCP
LLMs reason by analogy, copying industry norms. This engine is a 6-pivot cognitive trap that forces the agent to discard jargon and derive original solutions exclusively from physical, mathematical, or logical axioms.