Server exposes 6 security-focused tools via FastMCP with HTTP transport. All tools have descriptions (10-120 chars) and visible input schemas with typed parameters. However, descriptions are terse and lack actionable context for LLM selection. Parameter descriptions are minimal (5-40 chars). No output schemas documented. Error handling is absent, tools return raw JSON from backend without recovery guidance. Tools follow verb_noun naming but lack specificity (e.g., 'security_scan' vs 'scan_for_prompt_injection'). No tool annotations (readOnlyHint, destructiveHint). Composition is sound, each tool has one responsibility, but chaining context is missing.

• Add explicit output schemas to all tools. Document the JSON structure returned by the backend (e.g., security_scan returns {threats: [{type, severity, location}], risk_level: string}). Include field types and whether fields are always present.
• Expand parameter descriptions to 50-150 chars. For 'sensitivity', write: 'Sensitivity level (low: minimal checks, medium: standard checks, high: strict checks, critical: maximum checks). Affects false-positive rate.' For 'content_type', list valid values: 'prompt, text, code, message'.
• Add readOnlyHint: true to all tool definitions in mcp_server.py. Use @mcp.tool(annotations={'readOnlyHint': True}) or equivalent FastMCP syntax to signal that tools are safe to retry.
• Implement error recovery guidance. Wrap backend responses in a helper that transforms errors into actionable messages. E.g., if agent_id is not found, return: 'Agent ID not found. Verify the agent_id parameter or call identity_scope_check with a known agent_id first.'
• Enhance tool descriptions to 80-150 chars. E.g., security_scan: 'Scan user input for prompt injection, jailbreak attempts, PII leakage, and data exfiltration. Call before passing untrusted content to downstream tools. Returns threat list and risk level (low|medium|high|critical).' This guides LLM selection and sets expectations.