Loading ToolBench
An MCP server over the Cognitive Vulnerability and Mitigation Catalog (CVMC v0.1). Agents and platforms call it to find out how an interaction is likely to push people apart, and what reduces it.
Seam demonstrates solid tool design with clear, domain-specific naming and comprehensive descriptions. All four tools follow verb_noun patterns (check_, list_, get_) and include 150-250 character descriptions that explain WHAT, WHEN, and WHY. Input schemas are present and typed with Zod. However, output schemas are not formally documented, responses are JSON-serialized but lack explicit field definitions. Error handling is minimal (get_entry returns isError flag but lacks recovery guidance). No tool annotations (readOnlyHint, idempotentHint) despite all tools being read-only. Parameter descriptions are strong but could benefit from format constraints and examples.
Call this before producing, ranking, or sending content that a person will read: a reply in a contested thread, a ranked feed of civic information, or a summary of a disputed claim. Returns the cognitive vulnerability patterns in play, the mitigation for each, the measurement axis that tests whether the mitigation worked, a burden-and-equity check, and the governance hook. Use it to learn how content is likely to affect the reader's capacity to reason. Accuracy checking is a separate question this tool does not answer.
Retrieve one catalog entry with every field the schema defines: the AI affordance, the cognitive capacity it threatens, the failure pattern, the candidate intervention, non-friction safeguards, the evaluation metric, the burden-and-equity check, and the governance hook. Named patterns return the fields that exist and say so.
Return the eight fields every catalog entry must populate, the six measurement axes, the five cognitive capacities, and the catalog's provenance and caveats. Call this to understand what a catalog entry commits to before relying on one, or to propose a new entry in the same shape.
List the CVMC v0.1 patterns, optionally filtered by the cognitive capacity they threaten, by whether they are developed against the full eight-field schema, or by whether they bear on civic discourse. Use this to see what the catalog covers before checking an interaction against it.
Output schemas not formally documented. Tools return JSON via json() helper but lack explicit field definitions, type info, or structure guarantees. LLMs cannot reliably extract fields or plan downstream calls.
No tool annotations despite all tools being read-only. Missing readOnlyHint metadata prevents clients from optimizing caching, retry logic, and safety guarantees.
Error handling lacks recovery guidance. get_entry returns isError + bare message 'No entry X. Call list_patterns for valid ids.' but does not suggest alternatives or enable self-correction.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
check_interaction fallback to lexical matching on assisted-match failure is silent and undocumented. LLM receives caveat field but tool description does not explain this two-tier matching strategy.
list_patterns accepts optional filters but does not document pagination or result limits. No mention of max result count or whether results are capped.