MCP server for Purify's auditable security-intelligence feeds (CISA KEV / EPSS / enriched)
Four well-named tools with strong descriptions (194 - 392 chars, well above baseline 194 avg). All tools have explicit input schemas with typed parameters and clear field descriptions. Tool names follow verb_noun pattern (list_feeds, search_feed, get_provenance, check_feed_changed). Schemas are complete with enums for operators and feed names. However, output schemas are not formally documented in the code, responses are inferred from implementation. Error handling is present but lacks recovery guidance (e.g., 'unknown feed' error could suggest available feeds). No parameter descriptions for some nested filter fields. Security is solid (read-only operations, no secrets in params). Composition is clean, each tool has one responsibility.
Check whether a feed has changed since a given manifest hash. Returns true if the feed's current artifact hash differs from the one you pass; false if identical. Useful for incremental consumers: poll this before re-fetching a large feed.
Return the data passport of records matching a cve_id in a feed: the per-record _purify block (source, fetch time, extractor, raw artifact hash/URI) plus the feed's run chain (spec, run id, dataset artifact hash). Use this to cite or audit any answer derived from a feed.
List the available Purify security-intelligence feeds with their current version (dataset artifact hash), record count, last run time and fields. Feeds are continuously scheduled, quality-gated and self-healing; every record carries full provenance.
Query one Purify security feed. Filters combine with AND; ops: ==, !=, >=, >, <=, <, contains. epss/percentile compare numerically, dates as ISO strings. Returns JSON with the feed version (artifact hash), exact total_matched, and up to 50 records (sort with sort_by/order, trim with fields/limit). Use total_matched for counting; narrow filters when a result is capped. Feeds: 'kev' (CISA Known Exploited Vulnerabilities), 'enriched' (EPSS+KEV joined), 'epss' (EPSS high-risk).
Output schemas not formally documented. Responses are inferred from code (manifest struct, record arrays, boolean flags). LLMs cannot plan downstream operations without knowing return field names and types.
Error messages lack recovery guidance. 'unknown feed' error does not suggest available feeds; 'API error' does not indicate whether to retry. Errors should guide LLM to next action.
Nested filter object parameters lack individual field descriptions. The 'filter' array items have 'field', 'op', 'value' but descriptions are only in the parent. LLMs benefit from per-field guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | 2026-07-28+ | v2 |
search_feed result limit (50 rows) is enforced in code but not explicitly stated in tool description. Users/LLMs should know upfront that results are capped.
No pagination parameters (offset/limit) in list_feeds. If feed record counts grow large, response could bloat context. Consider adding optional limit/offset.