Loading ToolBench
MCP Server that executes TypeScript code in a sandbox with access to MCP tools (filesystem, fetch, etc.). Designed to work with Claude Code by providing code execution capabilities without requiring an ANTHROPIC_API_KEY.
This server has 14 tools with basic schemas and descriptions, but significant quality gaps prevent higher scoring. Most tool descriptions are present but minimal (10-50 chars), falling below the 50-200 char baseline for LLM-optimized descriptions. Parameter descriptions are sparse or missing entirely, many parameters lack any guidance on format, constraints, or valid values. The execute_code tool description is in Chinese and overly long (400+ chars). No output schemas are documented, forcing LLMs to guess what fields to expect. Error handling is minimal, no recovery guidance, no categorization of retryable vs fatal errors. The server lacks tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite having clear read/write semantics. Composition is reasonable (tools are single-purpose), but parameter naming could be more explicit (e.g., 'path' alone vs 'file_path').
Create new directory or ensure it exists
Edit file with diff-based changes
在安全沙箱中执行 TypeScript 代码。代码可以调用已连接的 MCP servers（filesystem、fetch 等）。 使用方式： 1. 编写调用 MCP 工具的代码 2. 代码在沙箱中执行 3. 返回 console.log 的输出 示例： import * as fs from "./servers/filesystem/index.js"; const content = await fs.readFile({ path: "package.json" }); console.log(JSON.parse(content));
Fetch URL and extract contents as markdown
获取 MCP 服务器连接状态和错误信息
Get metadata about file or directory
List directories available to access
No output schemas documented. LLMs cannot predict response structure, forcing them to guess what fields are available for downstream tool calls. This breaks tool chaining and wastes tokens on discovery.
Parameter descriptions are missing or minimal. 'path', 'edits', 'pattern', 'max_length', 'start_index', 'raw' lack guidance on format, constraints, or valid values. LLMs cannot infer whether 'path' is absolute/relative, whether 'pattern' is regex/glob, or what 'edits' structure looks like.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 49 | 2026-07-28+ | v2 |
列出沙箱中可用的 MCP 工具（文件树结构）
List directory contents
Move or rename files and directories
Read complete contents of a file
Read multiple files simultaneously
Search for files matching pattern
Create new file or overwrite existing
Tool descriptions are too short (10-60 chars) and lack context. 'Read complete contents of a file' does not explain when to use read_file vs read_multiple_files, what happens if the file is binary, or what the response structure is. Baseline is 50-200 chars for LLM-optimized descriptions.
No error handling guidance. Tools do not document what errors are retryable, what user-fixable errors look like, or how to recover. E.g., write_file may fail due to permission denied, disk full, or invalid path, LLM has no guidance on next steps.
Missing tool annotations. Tools with clear read/write semantics (read_file, write_file, delete operations) lack readOnlyHint and destructiveHint annotations. This prevents clients from applying safety policies or warning users before destructive operations.