MCP server exposing a vendor/vulnerability registry parsed from legacy pipe-delimited files
Strong foundation with well-structured schemas, clear naming conventions, and comprehensive parameter descriptions. All 6 tools follow verb_noun patterns (list_, get_, count_). Input/output schemas are fully defined using Zod with proper type constraints. Tool descriptions are detailed and contextual (avg ~150 chars), guiding LLM selection. All parameters have descriptions and type constraints (enums, ranges, patterns). Tool annotations (readOnlyHint, idempotentHint) are correctly applied. Main gaps: output schemas lack explicit documentation in tool registration (only visible in separate schema file), error handling is minimal (no recovery guidance or actionable error messages), and no batch operations despite tools that agents might call in loops.
Count how many vulnerabilities match the given filters, without returning the records. Use this — not list_vulnerabilities — for any question that only asks for a number (e.g. 'how many critical vulnerabilities are open?').
Report data-loading quality statistics for the vendor/vulnerability source files — record counts and structural warnings from startup. Use for questions about data quality/completeness, not about specific records.
Get an aggregated risk summary for one vendor by id — severity breakdown, open/patched counts, average and max CVSS. Use for questions about a vendor's overall risk profile, not for listing individual CVEs.
Look up a single vulnerability by its internal id or CVE id (e.g. 'CVE-2021-44228'). Returns found:false if no match — that is a normal, non-error result.
List software vendors in the registry, optionally filtered by an exact category or by a case-insensitive name substring. Returns a paginated list — use this to browse or find vendors, not to look up vulnerabilities.
Output schemas defined in separate file but not explicitly linked in tool registration. LLMs cannot see outputSchema in registerTool() calls, only inputSchema. This violates pattern:tool requirement that output structure be documented.
Error handling returns generic error messages without recovery guidance. Example: 'No vendor found with id "X"' does not suggest next steps (e.g., 'Try list_vendors() to find valid IDs'). Violates pattern:recovery-guide.
get_data_health lacks output schema definition and has minimal description (inferred from context only). No explicit documentation of what fields the health report contains or when to call it.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2025-06-18+ | v2 |
Search and list individual vulnerability (CVE) records with filters (vendor, severity, status, CVSS, published date range, free-text search) and pagination. Do not use this just to count matches — use count_vulnerabilities instead, it is cheaper and returns a single number.
No batch operations for tools agents might call repeatedly (e.g., get_vulnerability for multiple CVE IDs). Agents will make sequential calls, wasting tokens and latency.
Date range validation (publishedFrom/publishedTo) returns generic error. Should include the specific constraint violated and suggest valid formats.