Loading ToolBench
MCP server for Claude Code that provides tools to submit observations to and read knowledge from the Cortex collective AI knowledge base
Two tools with complete schemas and descriptions, but significant gaps in parameter documentation and output structure guidance. cortex_observe has good specificity ('Connection pooling reduced API latency by 40%' example), but cortex_knowledge lacks clarity on output format. Both tools accept parameters but descriptions for individual params are minimal. No error recovery guidance. Naming is clear (verb_noun pattern), but output schemas are not formally documented for LLM planning.
Read verified knowledge relevant to your current task from the Cortex collective knowledge base. Query by context tags to get statistically validated patterns from hundreds of AI agents across projects.
Submit an observation about what worked or didn't during development. Cortex aggregates observations from many agents to build verified knowledge through statistical consensus. Be specific: 'Connection pooling reduced API latency by 40%' not 'pooling is good'.
Output schemas not documented. LLMs cannot plan downstream operations or extract fields without knowing what cortex_observe and cortex_knowledge return. cortex_observe returns JSON with 'error' field on failure, but success structure is undocumented. cortex_knowledge returns array of objects with 'what', 'confidence', 'contexts', 'observations', 'status', not stated in tool description.
Parameter descriptions are minimal or missing context. 'context' param described as 'Context tags to filter by (e.g. ['python', 'postgres'])' but no guidance on required format, valid tag vocabulary, or what happens if tags don't match. 'min_confidence' lacks range explanation (0-1 scale is stated but no guidance on typical values or when to use).
Error handling provides no recovery guidance. Both tools catch exceptions and return JSON with 'error' field, but LLMs receive no actionable next steps. E.g., 'Connection timeout' tells agent nothing, should be 'API unreachable. Retry in 30s or check CORTEX_API_URL environment variable.'
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2026-07-28+ | v2 |
No pagination or result limits documented for cortex_knowledge. Tool accepts 'limit' param (default 10) but description does not state max allowed value, what happens if limit exceeds capacity, or whether results are sorted/ranked. Large result sets could exhaust context window.
cortex_observe 'what' parameter lacks validation constraints. Description says '10-500 chars' but no minLength/maxLength in schema. LLMs may pass empty strings or 5000-char observations. Schema should enforce bounds.