RelayPlane MCP server has 10 tools with basic registration and Zod schema validation. However, critical gaps prevent higher scoring: (1) Tool descriptions are generic and lack LLM-optimized detail (e.g., 'Execute a single AI model call' does not explain WHEN to use vs relay_workflow_run). (2) Parameters lack type constraints and enums (e.g., 'language' in relay_exec_script accepts 'python', 'javascript', 'bash' but no enum declared; 'model' accepts freeform strings instead of enum of available models). (3) No output schemas are documented, LLMs cannot plan downstream calls. (4) relay_exec_script is an IRREVERSIBLE operation with no dry-run, confirmation, or rate-limiting mentioned. (5) Parameter descriptions are minimal (1 - 2 short phrases); LLM-optimized descriptions should be 50 - 200 chars explaining purpose and constraints. (6) Error handling is generic (switch/default case throws McpError); no recovery guidance or actionable error messages. (7) Composition issues: relay_run and relay_workflow_run overlap in purpose; unclear when to use each. (8) No pagination guidance for relay_runs_list (accepts limit/offset but no total_count or next_cursor documented). Average per-tool score: 42 across 10 tools.

• Expand tool descriptions to 50 - 200 characters. Example: relay_run → 'Send a prompt to a single AI model and receive structured output. Use this for one-off inference; for multi-step workflows with dependencies, use relay_workflow_run. Supports optional system prompt and JSON schema for output validation.'
• Add enums for constrained parameters. Example: language in relay_exec_script → enum: [python, javascript, bash, shell]; model in relay_run → query relay_models_list first or accept enum from SDK [openai:gpt-4o, openai:gpt-4-turbo, anthropic:claude-opus, ...].
• Document all output schemas in tool definitions. Add a 'returns' or 'output' section listing fields, types, and examples. Example: relay_models_list returns [{model_id: string, provider: string, pricing: {input_per_token: number, output_per_token: number}, capabilities: string[]}].
• Add pagination documentation to relay_runs_list. Include 'total_count' or 'next_cursor' in response schema; state default and max limit (e.g. default limit=20, max=100); explain how to fetch all results.
• Enhance parameter descriptions with format constraints. Example: model → 'Model identifier in format provider:model_name (e.g., openai:gpt-4o, anthropic:claude-opus). Valid providers: [openai, anthropic, google, cohere]. See relay_models_list for available models.'
• Add structured error handling with recovery guidance. Catch validation errors and return: {error: 'Invalid parameter', param: 'model', value: '...', reason: 'Unknown provider', suggestion: 'Call relay_models_list() to see available models'}. For API errors, classify as retryable/user-fixable/fatal.