Loading ToolBench
A Python-based agentic RAG (Retrieval-Augmented Generation) system with tool ecosystem for document analysis, calculation, code execution, datetime operations, and file handling.
This server has 13 tools with moderate quality issues. Naming is generally good (verb_noun pattern mostly followed), but descriptions vary widely in completeness. Schemas are partially visible but lack formal JSON Schema type definitions in the code reviewed. Critical gaps: no error handling guidance, no output schema documentation, parameters missing type information in many cases, and no tool annotations. The code shows tool implementation but not proper MCP-style registration with complete schemas.
分析知识库文档的结构和组织方式，识别问题并提供优化建议。
进制转换工具，支持二进制、八进制、十进制、十六进制的相互转换。
安全数学计算器，支持加减乘除、幂运算、常用数学函数(sqrt, sin, cos, log等)。支持中文运算符(×÷)。用于精确数学计算。
在安全沙箱中执行代码并返回结果。 支持的语言: Python, JavaScript, Shell 适用场景: - 数据分析和计算 - 生成图表和可视化 - 文本处理和转换 - 快速验证代码逻辑 注意: 代码在隔离环境中执行，无法访问网络和敏感系统资源。
获取当前日期和时间，支持不同时区。可用于回答'现在几点'、'今天星期几'等问题。
对数据进行统计分析。支持: - 基本统计（均值、中位数、标准差等） - 数据转换（排序、过滤、分组） - 简单可视化描述 输入可以是 JSON 数组或 CSV 格式的数据。
日期计算工具：计算两个日期之间的天数差、在某个日期上加减天数、计算距离某个节日或截止日期还有多少天。
列出指定目录中的文件和子目录。
No formal JSON Schema type definitions visible in tool parameter registration. Parameters declared in docstrings/arrays but lack explicit 'type' fields in schema format (e.g., 'analysis_type' and 'style' lack type:string enforcement).
No error handling guidance or recovery suggestions in tool descriptions. Tools like code_executor and write_file (which have destructive/irreversible consequences) do not document what to do on failure or how to retry.
No output schemas documented. Tool descriptions do not specify what fields/structure the LLM should expect in responses, forcing the agent to infer output format.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 44 | 2026-07-28+ | v2 |
读取指定文件的内容。支持文本文件（.txt, .md, .py, .json 等）。
总结长文本或多个文档的内容，生成简洁的摘要。
单位换算工具，支持长度、重量、温度、面积、数据存储、时间、速度等单位换算。支持中文单位(如 公里、千克、摄氏度)。
世界时钟，同时显示多个城市/时区的当前时间。用于比较不同地区的时间。
将内容写入文件。如果文件存在会覆盖，不存在则创建。⚠️ 请谨慎使用，仅限允许路径。
Tool annotations (readOnlyHint, destructiveHint, idempotentHint) missing entirely. The server provides risk tags in the summary ('READ_ONLY', 'WRITE', 'IRREVERSIBLE') but these are not exposed via the MCP protocol.
Some parameter descriptions reference example values that LLMs may reuse literally (e.g., 'like 2025-01-01', 'like 北京,纽约,伦敦,东京'). This invites hallucinated values.
Parameters with constrained values (e.g., 'analysis_type', 'language', 'operation') documented as free-text with examples rather than declared as enums.
code_executor tool description notes it runs 'in a safe sandbox' but does not document what resources ARE accessible, what exceptions may occur, timeout behavior, or how the LLM should handle execution failures.
write_file lacks a dry-run or confirmation pattern. Agents could accidentally overwrite files without warning. No mention of append mode dangers or path traversal validation.
read_file accepts 'file_path' as relative path but does not document path sanitization or traversal guards. No limits on file size or what file types are readable.
list_directory with 'recursive=true' could return enormous results with no pagination or limit. No documentation on max depth or result cap.