Loading ToolBench
This server has critical definition quality gaps that would prevent reliable LLM agent usage. Tool naming lacks clear action verbs, descriptions are present but insufficient for agent planning, input schemas are incomplete (parameters lack type annotations in the visible schema definition), and output schemas are undocumented. The server exposes API keys as environment variables (good), but the parameter documentation is minimal. Both tools show evidence of the schema being inferred from code rather than explicitly visible in the registration, which caps per-tool scores.
该工具使用大语言模型来生成 HTML，并弹出 WebView 窗口，以此获取用户输入。参数为需要生成的页面的自然语言描述，包括标题、需要输入的内容、交互形式等。需要言简意赅但准确完整。
当调用 create-ask-ui 工具后，用户没有立刻给出结果，则可以调用该工具来再次尝试获取结果。该工具可以连续多次调用，直到用户给出结果。
Tool names lack clear action verbs. 'ask-ui' and 'continue-ask-ui' do not follow verb_noun naming convention (e.g. create_ui, fetch_result). LLMs cannot easily parse intent from the name alone. Should be 'create_ui_window' or 'show_ui_dialog' and 'poll_ui_result' or 'fetch_ui_response'.
Input schema for 'ask-ui' lacks explicit type annotations in the Zod definition. The code shows z.string() and z.number() but the schema object definition visible in registration does not include explicit 'type' fields per JSON Schema. Parameters 'prompt', 'html', 'title', 'height' descriptions are terse (10-20 chars).
Input schema for 'continue-ask-ui' is empty ({}). Zod schema is present in code but intentionally empty. This tool has zero input parameters, making it a pure stateful operation. No schema documentation present.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 24 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 17 | - | v1 |
No output schemas are documented. Both tools return text content but the structure of the response (what fields it contains, what data types) is not formally declared. LLMs cannot plan downstream operations without knowing return structure.
Tool descriptions are in Chinese. The 'ask-ui' description is in Mandarin: '该工具使用大语言模型来生成 HTML，并弹出 WebView 窗口...' LLMs trained primarily on English data may have reduced comprehension. Additionally, the description is functional but sparse (120 chars) and does not explain when to use this tool vs alternatives, what parameters do, or when it fails.
Parameter 'height' has no description at all in the visible schema registration. The 'title' parameter has a generic description ('Window title'). 'prompt' and 'html' descriptions are minimal (20-50 chars) and do not explain the fallback logic (when to use 'prompt' vs 'html', what happens if both are set).
Mutual exclusivity between 'prompt' and 'html' parameters is enforced at runtime (line 'if (html && prompt) throw new Error(...)') but not documented in parameter descriptions. LLMs will not know they are mutually exclusive and may pass both, resulting in an error.
Error messages are present but not recovery-guided. E.g. 'Last session is not finished' and 'Session not found' do not tell the LLM what action to take next. Should include guidance like 'Call continue-ask-ui to fetch the pending result' or 'Create a new session with ask-ui first'.