Loading ToolBench
Local MCP server for designing and scoring practical AI-agent evaluations.
Three tools with complete schemas and descriptions, but significant gaps in parameter documentation and output schema clarity. Tool names follow verb_noun convention (generate_, score_, evaluation_). Descriptions are present (avg 120 chars) but lack actionable context for LLM selection. All tools declare readOnlyHint and idempotentHint annotations. Critical issue: output schemas are not documented, LLMs cannot infer what fields to expect from responses. Parameter descriptions exist but are minimal (e.g., 'Task description' lacks format/constraint guidance). No error handling guidance visible.
Return a reusable scorecard structure for documenting AI-agent tests.
Generate practical test cases and success criteria for an AI-agent use case.
Create an unverified self-assessment of outcome, evidence, control and correction cost. Never use it alone to grant autonomy.
Output schemas not documented. Tools return JSON text but LLMs cannot infer field structure. generate_test_plan returns {useCase, risk, tools, cases, successCriteria} but this is not declared in tool definition. score_agent_run returns {total, dimensions, verdict, verification, reasons}, undocumented. evaluation_scorecard returns CSV header as text, unclear if this is the actual output format.
Parameter descriptions lack actionable constraints. 'useCase' described as 'Real task the agent must complete' but no guidance on format, length, or examples. 'risk' enum is declared but description does not explain when to use 'low' vs 'medium' vs 'high'. 'tools' array lacks guidance on tool naming conventions or expected format.
No error handling guidance. If validation fails (e.g., task field contains only whitespace after normalization), core.ts throws 'field doit contenir au moins un mot ou un nombre' (French error). LLM receives no recovery hint, should suggest 'Provide a task with at least one word or number' and optionally offer examples.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
evaluation_scorecard has empty input schema (no parameters) but description is vague ('Return a reusable scorecard structure'). Does not explain what the scorecard contains, when to call it, or how to use the output. Baseline: discovery tools should explain what structure they reveal.
Regex pattern in meaningfulText schema ([A-Za-zÀ-ÖØ-öø-ÿ0-9]) is not documented in parameter descriptions. LLMs cannot read JSON Schema patterns, they rely on description text. Users/LLMs will not know that special characters or punctuation are rejected.