Loading ToolBench
A Model Context Protocol server that enforces language and reasoning rules for Claude interactions. It monitors thinking blocks for language compliance (Korean vs English/Japanese), validates tool usage patterns, and manages session state with denial tracking and pattern matching.
Nekoi MCP provides 4 read-only tools for monitoring censorship/thinking patterns. All tools have descriptions (10-200 chars, within baseline range) and basic input schemas. However, schemas lack proper type definitions for most parameters, only 'limit' and 'text' have explicit types. Parameter descriptions are minimal (5-15 chars), well below the 72-char baseline. No output schemas are documented. Tool names use verb_noun pattern (censor_*) but are domain-specific and lack clarity for general LLM audiences. No error handling guidance, no pagination metadata, no field-level documentation. STDIO transport caps protocol readiness at 50, further limiting overall score.
Classify text as Korean or English reasoning using the same rule the hook enforces, and report which banned thinking.json patterns it matches.
Report thinking blocks the watcher flagged as they were written, newest last. Covers reasoning that never reached a tool call and so was never seen by a hook.
Report the censor's tally for one session: English/Japanese thinking counts, denials, tool calls, and current streak. Omit session_id for the most recently active session.
List supervised sessions, newest first, with each one's violation tally.
Input schemas lack type definitions. 'session_id' parameter in censor_session_status has no 'type' field; only 'limit' (integer) and 'text' (string) are properly typed.
Parameter descriptions are minimal (5-15 chars). Baseline is 72 chars. E.g., 'session_id' → 'Session id; defaults to the most recent.' lacks format, range, or dependency hints. LLMs cannot infer valid values.
No output schemas documented. Tools return structured data (tally, sessions list, thinking blocks) but LLMs have no schema to plan downstream calls or extract fields. Baseline: 100% of A+ tools document return types.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 48 | <=2025-11-25 | v2 |
No error handling guidance. Tools provide no recovery hints (e.g., 'Session not found. Call censor_sessions() to list available sessions.'). Baseline: error responses must guide LLM's next step.
STDIO transport only. Not remotely accessible; cannot be used by hosted MCP clients.