Static source inference · medium confidence · detected: Logging
Deprecated protocol patterns detected
Summary
This MCP server has severe definition quality issues. Only 2 tools are defined, and both lack critical documentation. The capture_screen tool has a minimal description (28 chars) and a parameter (save_path) with only a brief note about being 'optional'. The execute_query tool has a slightly better description (52 chars) but parameters lack actionable guidance on format/constraints. Neither tool has documented return schemas. The code shows working implementations (pyautogui for screen capture, DuckDB for CSV queries), but the tool definitions themselves are sparse and would not guide an LLM effectively through tool selection or parameter binding. Error handling descriptions are absent, LLMs cannot determine recovery paths on failure.
Tool descriptions are too short (28 chars for capture_screen, 52 chars for execute_query). Baseline for A+ tools is 50-200 chars. Descriptions lack context for when/why to use each tool.
Parameter descriptions are minimal or absent. 'save_path' in capture_screen only states 'Optional custom save path', no format guidance, no default behavior if omitted. 'csv_file_path' lists comma-separated paths but provides no examples or constraints (max file count, path format, allowed extensions).
No error handling guidance. If capture_screen fails (permission denied, no screen attached), or execute_query fails (invalid SQL, missing file), the LLM has no recovery path.
Recommendations
Expand capture_screen description to 80-150 chars: 'Capture the current screen and save as an image. Returns the file path where the screenshot was saved. Optional custom save_path defaults to ./screenshots/. Useful for visual documentation, recording UI state, or passing visual content to the LLM for analysis.'
Expand execute_query description: 'Execute a SQL query against one or more CSV files using DuckDB. Automatically infers schema from CSV headers. Returns query results as a structured array with column names. Supports standard SQL syntax; see DuckDB docs for dialect specifics. Use for data analysis, aggregation, filtering.'
Add format guidance to capture_screen's save_path parameter: 'Optional absolute or relative directory path. Defaults to ./screenshots/. If omitted, creates screenshots/ in current working directory. Path must be writable; will raise PermissionError if not.'
Add detailed guidance to execute_query's csv_file_path: 'Comma-separated list of CSV file paths (absolute or relative). Max 10 files per query. Each file must exist and be readable. CSV headers become column names; types are inferred from first 100 rows. Example: "data/sales.csv,data/regions.csv"'
Add guidance to execute_query's query parameter: 'Valid DuckDB SQL query string. Supports SELECT, WHERE, GROUP BY, JOIN, aggregates, etc. See DuckDB dialect reference. Example: "SELECT region, SUM(sales) as total FROM read_csv(\'sales.csv\') GROUP BY region". Return value is an array of objects with query result rows.'
Spec posture evidence
Inferred effective spec: <=2025-11-25.
Relies on Logging (deprecated) - log to stderr or use OpenTelemetry
Parameter 'save_path' in capture_screen lacks constraints. No guidance on absolute vs relative paths, allowed directories, filename format, or what happens if the path is invalid. LLM may pass malformed paths.
execute_query accepts 'query' as free-form string with no format guidance, DuckDB SQL dialect hints, or examples. LLM may pass invalid SQL or unsupported syntax.
No pagination, result limits, or chunking for execute_query. If a CSV contains thousands of rows, the entire result set returns unfiltered, blowing context window and degrading LLM reasoning.
execute_query
Document return schema for capture_screen: 'Returns {"status": "success", "file_path": string, "width": int, "height": int, "format": "png"}' to enable downstream tools that accept image paths.
Document return schema for execute_query: 'Returns {"status": "success", "rows": [{...}], "columns": [string], "row_count": int}' so LLM knows how to navigate results.
Add error handling guidance: For capture_screen: 'If screen cannot be captured, returns {"error": "CAPTURE_FAILED", "reason": "..."}. Possible reasons: PermissionError (no display), OSError (unsupported platform). Retry with explicit save_path or check display availability.' For execute_query: 'If SQL is invalid, returns {"error": "SYNTAX_ERROR", "detail": "..."}. If CSV file missing, returns {"error": "FILE_NOT_FOUND", "path": "...", "suggestion": "Check file path and permissions."}.'
Add result pagination to execute_query if large result sets are expected: Accept optional 'limit' (default 100) and 'offset' (default 0) parameters. Return total_row_count so LLM can fetch additional pages if needed.
Add idempotent hint: Both tools should include 'idempotentHint: true' in their definitions since re-running with the same inputs produces the same output (no side effects beyond file writes).