Loading ToolBench
MCP server enabling real-time user intervention in AI-assisted development workflows through Web UI feedback collection.
Single tool 'interactive_feedback' has a complete input schema with all parameters typed and described. However, the tool description is in Chinese and exceeds optimal length (400+ chars vs. baseline 194 chars). Output schema is not documented. The tool performs a single, well-defined action (requesting user feedback via Web UI), but lacks error handling guidance, recovery patterns, and confirmation mechanisms for a high-latency blocking operation. No parameter enums despite accepting predefined_options. The schema is visible and complete, raising the baseline, but missing output documentation and error patterns prevents a higher score.
本 MCP 服务暴露唯一工具 `interactive_feedback`，用于通过 Web UI 向人类用户请求澄清、决策或签收。开放世界工具（与真人交互，可能推送通知）。非破坏性（不会修改源代码 / git / 数据库）。非幂等（每次调用都会创建一个新的反馈任务）。阻塞直到用户提交、自动重调倒计时触发或后端超时。
Output schema not documented. LLM cannot plan what data will be returned (user response, timeout status, etc.) and cannot route output to downstream tools.
Tool description is 400+ characters (baseline 194), in Chinese, and lacks English clarity on WHAT the tool does vs. WHEN to use it. LLMs benefit from 50-200 char descriptions stating intent explicitly.
No error handling guidance. Tool blocks until user submits or timeout, but LLM has no recovery instructions if timeout fires, if user closes the panel, or if backend fails. Should document: retryable vs. fatal errors, next steps on timeout.
No confirmation or dry-run pattern. Tool is non-idempotent (each call creates a new feedback task) and blocks indefinitely if infinite_wait=true. High-latency operations risk agent hanging. Should document retry behavior and timeout defaults clearly.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 59 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 0 | - | v1 |
predefined_options parameter accepts 'array' but does not specify enum constraints or item schema (string vs. object with label/default). LLM cannot validate which options are valid without trial-and-error.
Tool name 'interactive_feedback' is vague about direction and action. Alternatives like 'request_human_feedback', 'wait_for_user_decision', or 'get_user_approval' would signal the blocking, human-in-the-loop nature more clearly.