An MCP server that dynamically exposes FastAPI endpoints as MCP tools by parsing OpenAPI specs
Critical gaps across all dimensions. Tool 1 (list_api_endpoints) has a generic description and empty input schema. Tool 2 (dynamic_tools) accepts a single 'params' string parameter with minimal schema detail, forcing LLMs to construct JSON manually, error-prone and violates constrained-input patterns. No parameter descriptions beyond the single 'params' field. No output schemas documented. No error handling guidance. The dynamic tool registration pattern obscures individual tool definitions from static analysis, making it impossible to verify per-endpoint schemas, descriptions, or error recovery paths. No tool annotations (readOnlyHint, destructiveHint, idempotentHint). No pagination support visible. Descriptions lack WHEN/WHY context and actionable recovery guidance.
Dynamically generated tools from OpenAPI endpoints. Each endpoint from the target FastAPI application is registered as a separate MCP tool with the pattern: {method}_{path_segments}. For example, GET /tasks becomes get_tasks, POST /tasks/{task_id} becomes post_tasks_task_id. Each tool accepts a JSON string parameter containing path parameters, query parameters, and request body fields.
Lists all available API endpoints that have been registered as MCP tools. Use this to discover what operations are available before calling them.
No input schemas for list_api_endpoints; dynamic_tools accepts only a single 'params' string with no type constraints, enum values, or format guidance. LLMs must construct raw JSON, inviting malformed payloads.
No output schemas documented for either tool. LLMs cannot predict response structure, making downstream tool chaining impossible and forcing agents to parse unstructured text.
Dynamic tool registration obscures individual endpoint definitions from static analysis. Tool names, descriptions, and schemas are generated at runtime and not visible in source code, preventing verification of per-endpoint quality and making it impossible to audit tool composition.
No error handling guidance. execute_request() catches exceptions but returns raw strings without categorizing errors as retryable, user-fixable, or fatal. LLMs receive no recovery hints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 36 | 2026-07-28+ | v2 |
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Agents cannot distinguish safe read-only tools from destructive operations, risking accidental data loss.