Server has 6 tools with consistent naming (verb_noun pattern: get-*) and basic Zod schemas. All tools have descriptions (avg ~60 chars, below 194-char baseline). Schemas present but lack output documentation. Parameters are mostly optional strings without enums or constraints. Error handling is minimal, failures return generic text responses without recovery guidance. No tool annotations (readOnlyHint, etc.). Parameter descriptions are present but lack format/range details (e.g., 'limit' says 'max: 5000' in description but no min/max in schema). No pagination metadata (total_count, next_cursor) documented. Responses are raw JSON dumps, not shaped for agent consumption.

• Add output schema documentation for each tool. Specify the structure of returned data (e.g., 'Returns object with fields: data (array of {id, symbol, name, price_usd, market_cap_usd}), status (string), timestamp (ISO 8601)').
• Convert free-form string parameters to enums. E.g., sort: enum(['market_cap', 'volume_24h', 'price']), sort_dir: enum(['asc', 'desc']), convert: enum(['USD', 'EUR', 'GBP', ...]).
• Add min/max constraints to numeric parameters in schema. E.g., limit: z.string().optional().refine(v => !v || (parseInt(v) >= 1 && parseInt(v) <= 5000), 'limit must be 1-5000').
• Implement structured error responses. Return {error: string, code: 'INVALID_PARAM'|'RATE_LIMITED'|'NOT_FOUND'|'API_ERROR', suggestion: string} instead of plain text. Guide LLM on retry strategy.
• Add readOnlyHint: true to all tool definitions via server.tool() options to signal safety to agents.