Loading ToolBench
MCP server for fetching web pages and extracting main content with ad removal and truncation
Single tool with complete JSON Schema and basic description. Schema is well-formed with proper types and defaults. Description is present but minimal (Chinese text, ~60 chars in English equivalent). Parameter descriptions are adequate but lack detail on constraints, error cases, and when to use the tool. No output schema documented. Error handling returns text responses but lacks recovery guidance or categorization. Tool is read-only and stateless, which is good, but lacks the depth expected for production use.
抓取网页并提取主体正文内容（去广告 + 截断）
Output schema not documented. Tool returns structured object with title, content, byline, siteName, url, truncated fields, but LLM has no formal schema to plan downstream operations or validate responses.
Tool description lacks context on WHEN to use it vs alternatives, prerequisites, and failure modes. Current description (Chinese) does not explain that it strips ads, truncates content, or requires valid URLs.
Parameter 'max_chars' lacks constraint documentation. No mention of minimum/maximum bounds, what happens if negative, or recommended ranges. Baseline rubric requires 'Specify minimum and maximum for numeric parameters'.
Error handling returns plain text error messages ('错误：...') without recovery guidance. Per pattern:recovery-guide, errors should tell LLM what to do next (e.g., 'Invalid URL format. Ensure it starts with http:// or https://').
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
No idempotency guarantee documented. Tool fetches live web content which may change between calls. LLM should know whether repeated calls with same URL return identical results or if content may differ.