Loading ToolBench
An MCP server indexed on Cloudflare's documentation at developers.cloudflare.com
Single-tool server with a semantically clear action verb ('search') and a moderately good description (100 chars, well within 10-1024 baseline). The tool name 'search_cloudflare_documentation' correctly signals intent. Input schema is present with a required 'query' parameter typed as string and described. Output is structured (XML text in content array). However, the tool lacks parameter constraints (no length limits, format guidance, or examples), error handling narratives, and response schema documentation. The description does not explain prerequisites, when to use this tool over alternatives, or what failure modes might require. Output is returned as unstructured XML text rather than richly typed fields, forcing the LLM to parse the response manually. No pagination support despite potentially large result sets (the code defaults to TOP_K=10 hardcoded). Schema is present but minimal, no per-field descriptions in the output structure, no guidance on result count or next steps. Does not meet A/B standards due to incomplete output schema documentation and lack of error recovery guidance.
Search Cloudflare's documentation and return semantically similar snippets
Output schema not documented. Returns XML-wrapped text in a single 'content' field. LLMs cannot infer the structure of result objects (url, text, similarity score) and must parse unstructured XML manually, increasing token waste and hallucination risk.
Input parameter 'query' lacks constraints. No length limits (min/max), format guidance, or examples. Baselines show production tools document length ranges and expected format for text params. LLMs may pass absurdly long queries or special characters without guidance.
No error handling narratives. If a query is malformed, the API is unavailable, or embeddings fail, the tool does not guide the LLM on recovery. 'Try again later' or 'simplify your query' guidance is absent.
Inferred effective spec: 2026-07-28+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 52 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 49 | - | v1 |
Tool description does not explain when to use this over other search tools (if any exist), what data is indexed, or prerequisites (e.g., valid Cloudflare API keys). It also does not mention the hardcoded 10-result limit, forcing the LLM to discover this behavior via trial.
No pagination support. Results are hardcoded to TOP_K=10. If the LLM needs more results, it cannot request them. Per baselines, tools returning lists should accept limit/page parameters and return total counts.