Loading ToolBench
MCP server that gives an MCP host a working sub-agent whose brain can be any Anthropic-compatible model
Single tool 'approval_prompt' has critical definition gaps. Description is 67 chars (acceptable length) but lacks actionable context for LLM selection. Input schema present but parameters lack descriptions, 'tool_name' and 'input' are defined with types but no guidance on format, constraints, or when to use them. No output schema documented. Tool name does not follow verb_noun pattern (should be 'request_approval' or 'get_approval'). No error handling guidance. Security concern: tool accepts arbitrary 'input' object without validation or sanitization, risking injection attacks.
Claude Code 用它询问某次工具调用是否放行。
Tool name 'approval_prompt' does not follow verb_noun pattern. Should be 'request_approval' or 'get_approval' to clearly signal the action to the LLM.
Input parameters 'tool_name' and 'input' lack descriptions. LLM cannot infer what format 'input' expects (JSON object? string? schema?). Parameter descriptions are mandatory for LLM reasoning.
No output schema documented. LLM cannot plan downstream actions or extract approval result (approved/denied/reason). All tools must document return types.
Tool accepts arbitrary 'input' object without validation or sanitization. Risks injection attacks if LLM passes malicious payloads. Validate and sanitize all agent-provided input.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 46 | 2026-07-28+ | v2 |
No error handling guidance. If approval is denied or tool_name is invalid, LLM receives no recovery path. Errors must tell the agent what to do next.