This is a test/demo server, not a production MCP implementation. Tool definitions are present with schemas and descriptions, but the server is intentionally incomplete: it does not actually execute tools, only fingerprints MCP endpoints. Descriptions are present but generic (avg ~100 chars). Schemas are minimal (only 'sql', 'to/subject/body', 'limit', 'path', no type info beyond string/integer). No parameter descriptions. No output schemas documented. No error handling guidance. The code comment explicitly states 'Nothing here is a real MCP implementation: it speaks exactly enough Streamable HTTP JSON-RPC to be fingerprinted.' This is a security scanner test fixture, not a usable agent tool.

• Add a description to every input parameter. E.g., 'sql: SQL SELECT statement (read-only, max 10000 chars, no DDL/DML)' and 'path: Relative file path in workspace (no ../ traversal, max 255 chars)'.
• Document output schemas for each tool. E.g., run_query returns {rows: [{col1: string, col2: number}], rowCount: integer}; send_email returns {messageId: string, timestamp: ISO8601}.
• Add enum constraints to parameters with known valid values. E.g., if 'limit' has a max, declare it as {type: integer, minimum: 1, maximum: 1000}.
• Expand tool descriptions to 50 - 200 chars explaining WHEN to use each tool and what happens on error. E.g., 'Send an email via the shared notifications mailbox. Use for alerts and confirmations. Fails if recipient is invalid; returns messageId on success.'
• Implement error classification: return {error: {code, message, category: 'retryable|user_fixable|fatal', hint: 'Try X instead'}} so LLMs know whether to retry, ask the user, or give up.