Loading ToolBench
OpenAI & Anthropic-compatible API gateway for Kiro (single-account, compliance-mode). Routes all completions through kiro-cli via ACP (Agent Client Protocol).
Single tool (web_search) with basic schema and description. Tool name lacks action verb prefix (should be 'search_web' or 'web_search' is acceptable but borderline). Description is present (67 chars) but minimal, lacks WHEN to use, prerequisites, or return structure guidance. Input schema has types and descriptions for both params (query: string, max_results: integer), meeting baseline. However, no output schema is documented, LLMs cannot predict what fields to expect from results. Error handling exists (MCPToolError) but lacks recovery guidance. No parameter constraints (max_results unbounded). Tool is read-only (good), but lacks idempotent/readonly annotations. Overall: functional but below production baseline (194 char avg description, 100% A+ tools document output).
Call a native web-search MCP tool and return a formatted summary.
No output schema documented. LLMs cannot predict result structure (fields, types, pagination). Baseline: 100% of A+ tools document return types.
Tool description (67 chars) lacks WHEN to use, prerequisites, and return structure. Baseline: avg 194 chars, 94% of A+ tools have complete docstrings.
max_results parameter unbounded. No min/max constraints documented. Baseline: numeric params should specify range (e.g., 1 - 100).
Error handling returns raw exception text ('Web search failed: {exc}') without recovery guidance. Baseline: errors should guide LLM on next action.
No tool annotations (readOnlyHint, idempotentHint). Tool is read-only but LLM cannot infer this from schema alone.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | <=2025-11-25 | v2 |