Demo of MCP OAuth 2.1 integration with AWS Cognito — showcasing pre-registered, DCR, and CIMD client registration methods
This server exhibits severe definition quality gaps across nearly all dimensions. The six tools have minimal to no descriptions, completely lack input/output schemas, and lack proper error handling. The tool names are HTTP paths rather than action verbs (GET /.well-known/..., POST /oauth/token), which violates fundamental naming conventions. Most critically, three of six tools claim to have input schemas in the evaluation metadata, but the actual source code provided does NOT show explicit tool registration, schema definitions, or MCP server setup. The server appears to be a raw Express HTTP server proxying OAuth flows, not a properly instrumented MCP server with tool definitions. This is a fundamental architectural mismatch, an HTTP API is not the same as an MCP server with properly defined tools.
Generic OAuth authorization server metadata endpoint that proxies to AWS Cognito and augments metadata with registration_endpoint, PKCE support, and CIMD support indicators
Protected Resource Metadata endpoint that returns OAuth resource metadata including authorization servers and supported scopes
OAuth Authorization Proxy that transparently handles URL-based client_ids (CIMD) by resolving them to Cognito credentials, then redirects to Cognito's authorization endpoint. Non-URL client_ids are passed through unchanged.
MCP API endpoint to fetch contexts with OAuth token validation
OAuth Token Proxy that transparently handles URL-based client_ids (CIMD) by mapping them to Cognito credentials, then forwards the token request to Cognito. Non-URL client_ids are passed through unchanged.
MCP API endpoint to create a new context with OAuth token validation
No MCP tool registration visible in source code. The provided code shows only an Express HTTP server with OAuth endpoints, not an MCP server with registered tools, schemas, or descriptions. Tool names, descriptions, and schemas in the evaluation metadata cannot be verified in src/mcp-server/index.js.
Tool naming violates action-verb convention. Names like 'GET /.well-known/oauth-protected-resource' and 'POST /oauth/token' are HTTP paths, not action verbs. MCP tools should use verb_noun naming: 'get_oauth_metadata', 'exchange_oauth_token', 'list_contexts', 'create_context'. HTTP paths as tool names confuse LLMs about the tool's purpose.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 27 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 16 | 2025-06-18+ | v1 |
All tools lack input/output schemas. The evaluation metadata claims schemas exist for tools 3, 4, and 6, but the source code does not show explicit schema definitions. Even for tools with claimed input schemas, output schemas are completely missing. LLMs cannot infer required parameters or expected response structure.
Descriptions are either missing or generic. Most tools have descriptions under 50 characters or lacking actionable detail. For example, 'Protected Resource Metadata endpoint that returns OAuth resource metadata...' (65 chars) lacks context on WHEN to use this tool or what the response structure is. Descriptions should be 50-200 characters with WHAT, WHEN, and RETURN information.
No error handling or recovery guidance. There is no indication of how tools handle errors, what codes are returned, or what an LLM should do if a call fails. No retryable/user-fixable/fatal classification. This makes agent error recovery impossible.
Parameter descriptions for tools with claimed schemas are generic or missing. For example, POST /oauth/token lists 'client_id' with description 'OAuth client identifier', this lacks format constraints, whether it's required, and how it relates to CIMD (Client ID Metadata Document) resolution described in the tool description.
Secrets exposure risk. POST /oauth/token accepts 'client_secret' as an input parameter. Credentials must never appear as tool parameters, they should be injected server-side via environment variables or vault. Agent logs will expose this secret.
No pagination or result limiting documented. GET /v1/contexts has no limit or offset parameters and no indication of how many results are returned. Without pagination, large result sets will exhaust the context window.