The Security Scanner MCP server has three tools with adequate naming but inconsistent quality across definitions. Tool names follow verb_noun convention (scan_for_sensitive_data, redact_text, validate_data_safety) which is positive, but descriptions lack depth and schema documentation is minimal. Parameters have type information but lack detailed constraint descriptions. Error handling exists but provides generic messages rather than actionable recovery guidance. No parameter descriptions in the schema itself, only in docstrings. Output structure is present but not formally documented in schema. The server falls into the 'fair to poor' category typical of community MCP implementations.

• Add formal output schemas to each tool. Document the exact structure: e.g., scan_for_sensitive_data returns {has_sensitive_data: boolean, sensitive_data_types: string[], findings_count: number, redacted_content: string, findings_detail?: {[key: string]: string[]}, alert: string}. This lets LLMs plan downstream extraction.
• Expand parameter descriptions in the @mcp.tool() registration. For 'allowed_types', replace short description with: 'List of allowed sensitive data types (e.g., password, api_key, token, email, phone, ssn, credit_card, ip_address, url_with_credentials). If empty, all detected types are disallowed.'
• Add tool selection guidance in descriptions. E.g., scan_for_sensitive_data: 'Use this to discover what sensitive data exists in text. Returns detailed findings and a redacted version. If you need to validate text against a policy, use validate_data_safety instead.' Validate_data_safety: 'Use this to enforce a data safety policy. Pass the allowed types (e.g., email, phone) and receive a boolean is_safe result. For discovery, use scan_for_sensitive_data.'
• Replace generic error handling with actionable messages. E.g., 'Error scanning content: input text exceeds 10MB limit. Try splitting into smaller chunks.' or 'Error parsing allowed_types: invalid value "creditcard", valid types are: password, api_key, token, email, phone, ssn, credit_card, ip_address, url_with_credentials.'
• Remove original_text from redact_text output. The LLM caller already has it. Return only: {redacted_text: string, redactions_made: string[], redaction_count: number, status: 'redacted'|'no_redaction_needed'|'error'}.