This OpenRouter MCP server exhibits strong definition quality across 19 tools with comprehensive descriptions, proper JSON Schema validation, and clear parameter documentation. All tools have non-empty descriptions (50-250 chars typical), input schemas with type definitions, and named parameters. However, there are notable gaps: (1) output schemas are not explicitly documented for any tool, forcing LLMs to infer response structures; (2) several tools lack actionable error guidance (e.g., what to do if a model validation fails); (3) some parameter descriptions reference external URLs without inline constraint documentation. The server demonstrates good adherence to verb_noun naming (chat_completion, analyze_image, generate_video) and includes helpful context like OpenRouter's provider-routing system and caching features. Transport is STDIO-only, which caps the overall protocol readiness score to 50 despite excellent definition quality.

• Add explicit output schema documentation to every tool. For chat_completion, document: { type: 'object', properties: { content: { type: 'string' }, tokens_used: { type: 'number' }, ... } }. For search_models, document the paginated structure with model objects.
• Enhance error handling with recovery guidance. E.g., validate_model: 'Returns true if model exists. If false, try search_models(query=partial_name) to find alternatives.' classify errors as retryable (transient network failure, rate limit) vs user-fixable (invalid model ID) vs fatal (auth failure).
• Move external references inline. For the `provider` parameter, inline the allowed keys: 'quantizations, ignore, sort, order, require_parameters, data_collection, allow_fallbacks' with type hints. Remove the external URL reference or keep it as a secondary note only.
• Clarify sandboxing and path validation behavior. Update analyze_image description: 'Local paths are sandboxed to OPENROUTER_INPUT_DIR. Absolute paths and traversal attempts (../etc/passwd) are rejected with a clear error.' Include this once per image/video/audio tool family rather than in parameter hints.
• Document caching semantics explicitly. E.g., 'cache: true enables response caching. Repeated identical requests return cached results if available (TTL: cache_ttl, default 1 hour). Cache misses trigger a fresh API call. Use cache_clear: true to invalidate and force refresh.'
• Add timeout and polling guidance to async tools. E.g., start_chat_completion: 'Returns job_id immediately. Poll with get_chat_completion_status() every 2-5 seconds. Expected latency: 10s (fast models) to 5m (reasoning models). Max retries: suggest 120 (10 min polling with 5s intervals) before reporting timeout to user.'

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry