MCP-Fresh has 13 tools with highly inconsistent quality. Basic calculator tools (add, subtract, multiply) have proper schemas and English descriptions, but the code analysis suite is severely underdocumented. Most critically: tools 4-6 and 9-13 have Chinese descriptions that are under 20 characters in English equivalent, making them non-compliant with description standards. No tool includes error recovery guidance, output schemas are undocumented, and parameter descriptions lack constraint information (e.g., file_path format, directory traversal safety). The codebase shows hardcoded absolute paths (/Users/yuanquan/code_project/GalaxyAI/MCP-Fresh) that compromise security and portability. Tools claim to be STDIO-based but the calculator_server.py and minimal_code_server.py files show uvicorn/FastAPI patterns inconsistent with standard STDIO MCP. Schema completeness varies widely, calculator tools have proper integer typing, but code analysis tools return complex nested dicts without documented structure.

• Translate all tool descriptions to English and meet 20-100 character minimum. E.g., 'count_lines' → 'Count non-empty, code, and comment lines in a file. Use for code metrics and statistics.'
• Remove hardcoded paths from tool implementations. Accept project_root as an optional parameter or read from environment variable MCP_FRESH_ROOT. Validate all file paths against a safe base directory to prevent traversal attacks.
• Document output schema for every tool as a JSON example or formal schema. E.g., count_lines returns {file_path: string, total_lines: int, non_empty_lines: int, code_lines: int, comment_lines: int}.
• Add error recovery guidance to all error responses. E.g., 'File not found: path. Try calling list_python_files(directory=".") to see available files.' Use pattern:recovery-guide.
• Add parameter constraints to descriptions. file_path: 'Path to a Python file (must exist, .py extension, no path traversal). Either absolute or relative to project root.' Provide examples of valid/invalid inputs.
• Clarify tool overlap: merge analyze_python() (basic AST) and analyze_code_intelligence() (semantic analysis) or split by clear intent. Rename one to avoid LLM confusion. E.g., 'analyze_python_ast' vs 'analyze_code_quality'.
• Add pagination to list_python_files() and get_project_overview(). Accept limit (default 20, max 100) and offset. Return {items: [], total_count: int, has_more: bool}.
• Implement per-request error categorization: retryable (network error), user-fixable (file not found), fatal (syntax error). Guide LLM on recovery action.