Probe has 4 tools with clear, action-oriented names (probe_scan, probe_scan_server, probe_list, probe_check_config) that follow verb_noun convention. Descriptions are present and substantive (100-200 chars), explaining what each tool does and when to use it. However, input schemas are minimal or absent: probe_scan and probe_list have empty input objects ({}), while probe_scan_server and probe_check_config have only one parameter each with basic type info. No output schemas are documented, LLMs cannot predict the structure of returned JSON reports. Error handling exists (e.g., 'Server not found' in probe_scan_server) but lacks recovery guidance. Overall, definitions are functional but lack the rigor expected of production tools.

• Document output schemas for all tools. Define the structure of FullReport, ScanResult, and Finding objects so LLMs know what fields to expect (e.g., 'Returns {results: [{server: {name, transport, command}, findings: [{severity, title, description, recommendation}], grade: string}], config_sources: [string]}')
• Add tool annotations to the FastMCP tool registration: @mcp.tool(readOnlyHint=True, idempotentHint=True) for all four tools. This enables agents to cache results and optimize retry logic.
• Enhance error messages with recovery guidance. Instead of 'Config file not found: {config_path}', return 'Config file not found: {config_path}. Try probe_list to see discovered config locations, or check that the path is absolute and readable.'
• Replace example values in parameter descriptions with formal constraints. Instead of 'e.g., "engram", "kv", "slack-bot"', document: 'Server name as it appears in MCP config (alphanumeric, hyphens, underscores; case-sensitive).'
• Add a 'format' or 'pattern' constraint to config_path parameter to validate absolute paths and reject relative paths early with a clear error message.