Loading ToolBench
A governance and compliance framework for enterprise AI systems with PII scrubbing, risk-tier gating, and provenance audit validation
This server exposes three tools via STDIO with minimal MCP integration. Tool definitions lack proper JSON Schema input specifications, parameters are documented only in docstrings, not in structured schemas. Descriptions exist but are domain-specific and lack LLM-optimization guidance (e.g., 'Entry-time PII scrubber' does not explain WHEN to call it or what happens on failure). No output schemas are documented. Tools are not registered with MCP tool definitions; they are standalone CLI scripts invoked via subprocess. This violates the core MCP pattern of tool registration with typed schemas.
Entry-time PII scrubber with two modes: scrub (removes PII patterns from JSONL events) and check (validates that events are clean of PII patterns)
Risk-tier gate that decides the gate for an action against a risk-tier policy by computing tier from blast_radius and data_sensitivity scores, applying irreversible floor, and matching explicit rules
Provenance audit-log validator that validates every record against the provenance schema's required fields and cross-field rules (R1-R5) ensuring delegation chains, reversibility constraints, and approval requirements are met
No JSON Schema input definitions. Tools are CLI scripts with argparse, not MCP tool definitions. Parameters lack type, enum, and constraint declarations required by pattern:constrained-input.
No output schemas documented. LLMs cannot infer what fields to expect from tool responses, violating pattern:tool and pattern:response-shaper.
Descriptions lack LLM-optimization. 'Entry-time PII scrubber' (27 chars) does not explain WHEN to use it, what happens on error, or what it returns. Baseline: 194 chars average for A+ tools.
Parameter descriptions embedded in docstrings, not in schema. LLMs cannot parse docstrings; they need structured parameter descriptions in the tool definition.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 24 | 2026-07-28+ | v2 |
No error handling guidance. Tools exit with code 1 on failure (e.g., pii_scrub --check), but no recovery hints provided to LLM. Violates pattern:recovery-guide.