mcp-prometheus demonstrates solid fundamentals across 15 well-named tools with consistent action-verb prefixes (execute_, get_, query_, check_). Descriptions are present and adequately detailed (mostly 60-150 chars, within production baseline of 34-392 chars). All tools follow a verb_noun naming convention. Input schemas are properly typed with JSON Schema format. However, several gaps prevent a higher score: (1) Output schemas are NOT documented in the provided source, only input schemas are visible. (2) Parameter descriptions lack specificity about format constraints, ranges, and examples of valid values (e.g., 'timeout' accepts strings like '30s' or '5m', but this is only hinted in description, not formalized as a pattern or enum). (3) No visible error handling guidance, tools do not document what errors to expect or how to recover. (4) Pagination/limiting strategy is inconsistent: 'execute_query' and 'execute_range_query' expose an 'unlimited' parameter with a performance warning, which is an anti-pattern, caps should be enforced server-side, not exposed to the LLM. (5) The 'org_id' parameter is repeated across all 15 tools, suggesting opportunity for server-level configuration rather than per-call injection. (6) Several tools ('get_config', 'get_flags') expose sensitive Prometheus internals without clear guidance on what the agent should do with that information.

• Document the complete output schema for every tool. Include field names, types, and whether lists are paginated. Example: 'Returns an object {data: [{metric_name: string, type: string, help: string}], total: integer, next_cursor?: string}'.
• Replace the 'unlimited' parameter with a server-enforced default limit (e.g., 1000 results max). Provide 'limit' and 'offset' parameters for pagination. Remove the performance warning from LLM-visible params.
• Formalize format constraints: convert 'timeout' description to explicitly state 'Go duration format (e.g., 30s, 1m, 5m)'. Convert 'start'/'end' to 'RFC3339 format (e.g., 2024-01-15T10:30:00Z)'. Use JSON Schema 'pattern' or 'format' fields, not just prose.
• Add error recovery guidance to each tool description. Example: 'If the query times out, retry with a larger timeout or simpler query. If no results are returned, check that the metric exists with get_metric_metadata().'
• Move 'org_id' from per-tool parameters to server configuration or MCP resource context. If multi-tenancy is required, make it a top-level client capability negotiated during initialization, not a per-call parameter.
• Add WHEN-to-use guidance to discovery tools. Example for 'get_metric_metadata': 'Call this first if you do not know what metrics are available. Results can be piped into execute_query().'
• Consider splitting 'get_series' and 'query_exemplars' into separate specialized tools or document when to prefer one over the other. Both query time-series data, the distinction should be explicit.

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

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