Loading ToolBench
Build custom MCP servers in minutes with 8 pre-built apps including web search, terminal, filesystem, and more
This MCP server has 30 tools spread across 6 modular applications (everything2md, mcp_factory, md_converter, rag_flow_mcp, rag_base, rag_eval_flow). Quality is highly inconsistent. Many tools have adequate descriptions and basic parameter schemas, but critical issues undermine overall reliability: (1) Tool naming lacks consistency, some use verb_noun convention (convert_to_markdown, create_dataset, delete_dataset) while others use compound names with poor clarity (mcp_rag_flow_fill_clarification_suggestions, mcp_rag_base_dataset_manage). (2) Parameter documentation is incomplete across most tools, many parameters lack descriptions entirely or have minimal context. (3) Output schemas are not documented anywhere in the provided code, critical for agent planning and chaining. (4) Error handling is generic and non-actionable ('Error: {str(e)}' with no recovery guidance). (5) Two tools (mcp_rag_base_dataset_manage, mcp_rag_base_document_manage) combine multiple unrelated operations under a single action parameter, violating single responsibility. (6) Security concerns: API keys and base URLs are loaded from config but the mechanism is not shown, cannot verify secret injection safety. (7) Tool descriptions in Chinese (everything2md, md_converter, rag_eval_flow) will confuse LLMs expecting English. The server shows decent foundational structure but lacks the polish and discipline required for production agent toolkits.
Atomic Tool: Apply suggestions to the document.
Build an existing MCP project into a standalone EXE with real-time progress.
提取 Markdown 文件中的表格并保存为 Excel (.xlsx) 文件。如果文件中有多个表格，将保存为不同的 Sheet。
将各种文档格式转换为 Markdown (异步)。支持格式: - Office: .docx, .doc, .xlsx, .xls, .pptx, .ppt - PDF: .pdf - Images (OCR): .png, .jpg, .jpeg, .tiff, .bmp
将 Markdown 文件转换为 PDF 文件。注意：需要系统安装有支持中文的字体 (如 SimHei)。
将 Markdown 文件转换为 Word (.docx) 文件。
Create a new Knowledge Base (Dataset).
Two mega-tools (mcp_rag_base_dataset_manage, mcp_rag_base_document_manage) combine CREATE/UPDATE/DELETE/LIST under a single action enum parameter, violating single-responsibility principle. This forces agents to reason about multiple operation types instead of selecting from discrete, focused tools.
Tool names prefixed with 'mcp_' and suffixed with lengthy operation descriptions (e.g., mcp_rag_flow_fill_clarification_suggestions, mcp_rag_flow_evolve_scheme_document) are cryptic and do not follow verb_noun convention. LLMs cannot easily distinguish these from context, names like 'fill_clarification_suggestions' and 'evolve_scheme_document' are clearer without the 'mcp_' prefix and would parse intent faster.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | <=2025-11-25 | v2 |
Atomic Tool: Create a shadow copy of the document (_ai_revision). Returns the path of the created shadow file.
Delete a Knowledge Base by ID.
Delete a document from a Knowledge Base.
评估回答质量 (Step 4)
Atomic Tool: Extract questions from a Markdown document (headers). Returns a list of identified questions with line numbers.
Scenario 1 Controller: Smart Clarification Suggestion Filling. Reads a Markdown file, identifies questions (Headers), retrieves answers from RAG, and fills them into a shadow copy of the file.
生成测试问答对 (Step 1 & 2)
Get parsed chunks of a document.
测试工具
Initialize a new MCP project with 6A workflow structure.
List all Knowledge Bases.
List documents in a Knowledge Base.
List all existing MCP projects in the factory. Returns a JSON string list of project names.
Manage Knowledge Bases (Datasets).
Manage Documents in a Knowledge Base.
[主线任务] 基于澄清决策进化方案文档。将已确认的澄清点应用到原方案文档中，生成 v1.1 版本。
[主线任务] 填充澄清建议 (Hybrid: Prefer Legacy Logic). 读取评审问题记录文档，调用 RAG 检索知识库，并将带有置信度的建议填入文档。
Atomic Tool: Retrieve a single suggestion from RAG. IMPORTANT: This tool expects a clean, well-formulated query. The Client/Agent should perform query rewriting (using its conversational LLM) BEFORE calling this tool if the original input is messy or ambiguous.
基于知识库执行问答测试 (Step 3)
Update Knowledge Base metadata.
Update document metadata (rename or enable/disable).
Upload a file to a Knowledge Base. Args: dataset_id: The target Knowledge Base ID. file_path: Absolute path to the local file.
Verify the built EXE of a project using smoke tests.
Output schemas are completely undocumented. The code shows function signatures and input schemas (parameters) but nowhere specifies what data is returned. Agents cannot plan downstream calls or extract fields without reverse-engineering from descriptions.
Error handling returns generic exception strings ('Error: {str(e)}') with no recovery guidance. When a tool fails, the LLM gets no information about whether to retry, fix input, or try an alternative tool. Example: 'Error uploading document: ConnectionError' tells the agent nothing actionable.
Tool descriptions are inconsistently in Chinese (everything2md, md_converter, rag_eval_flow tools) while most are in English. LLMs are trained on English documentation and descriptions, mixing languages degrades model behavior and increases parsing errors.
Many parameters lack descriptions entirely or have minimal context. For example, in mcp_rag_base_dataset_manage, the 'action' parameter description is 'One of [create, delete, update, list]' but does not explain when to use each or what fields are required per action. Similarly, 'name' parameter in create_dataset has no description of length constraints or allowed characters.
No output pagination documented. Tools like list_datasets, list_documents, and list_projects accept page/page_size, but return values are not visible, agents cannot know total count, whether more results exist, or how to chain pagination calls. This risks incomplete result sets being treated as exhaustive.
Destructive operations (delete_dataset, delete_document, delete_project) lack any confirmation mechanism or dry-run option. An agent could accidentally delete a critical knowledge base in one call with no recovery path. No tool description warns of irreversibility.
File path parameters (source_path, output_path, doc_path, file_path) lack validation rules in descriptions. No mention of absolute vs relative paths, allowed characters, or path traversal protection. Agents could pass malicious paths like '../../../etc/passwd' or Windows UNC paths, creating security vulnerabilities.
Tool composition is broken in several workflows. For example, apply_suggestions_to_doc expects a 'suggestions_map' as a JSON string, but retrieve_rag_suggestion returns suggestions in an unspecified format. Agents have no clear way to map individual RAG results into the JSON structure expected. This forces agents to reason about string formatting, risking malformed payloads.
No tool annotation hints (readOnlyHint, destructiveHint, idempotentHint) are present. Agents cannot determine at a glance whether a tool modifies state. For example, is 'update_dataset' idempotent? Can 'fill_clarification_suggestions' be safely retried? The lack of explicit annotations forces LLMs to infer from names and descriptions, risking wrong assumptions.