Prometeo MCP has 27 tools with significant quality gaps. All tools are explicitly registered via @mcp.tool() decorators and have descriptions. However, parameter descriptions rely heavily on a utility function (get_param_description) whose actual strings cannot be verified in the provided code, creating uncertainty about description quality. Schema definitions are present and well-typed using Pydantic annotations, but many lack detail in constraints and ranges. Error handling is basic, most tools return dict with status and message fields, but lack actionable recovery guidance. No tool annotations (readOnlyHint/destructiveHint/idempotentHint) are used, despite clear WRITE vs READ_ONLY distinctions in the metadata. Tool naming is clear and follows verb-noun patterns consistently, which is a strength. Composition is good, tools are single-responsibility and chainable. However, the reliance on an external description utility and STDIO-only transport are significant limitations.

• Audit and document all parameter descriptions from get_param_description utility. Verify they meet 10 - 1024 char guideline and clearly state WHAT the parameter controls, WHEN it's needed, and any constraints (format, range, enum).
• Add readOnlyHint, destructiveHint, and idempotentHint annotations to all tools based on existing Risk metadata (READ_ONLY, WRITE, REVERSIBLE). This enables LLM to reason about operation safety without reading descriptions.
• Enhance error messages with actionable recovery guidance. Instead of 'BankingClientError: invalid session', return 'Invalid or expired session. Call banking_login() with username and password to establish a new session.'
• Document output schemas for every tool. Use JSON Schema or prose to describe returned fields, types, and semantics. E.g., banking_login returns {status: 'success'|'interaction_required'|'select_client'|'error', session_key: string, message: string}.
• Refactor banking session management from global _active_sessions dict to per-request storage. Pass session_key as a parameter in each request; verify on the server that the session exists and belongs to the calling agent.
• Add enum constraints to country_code, bank_code, currency fields. Document allowed values in descriptions or use Pydantic Field(enum=[...]) to let LLM pick valid options.
• Add validation on banking_get_movements: (a) ensure start_date <= end_date; (b) cap date range to 90 days max; (c) document this in description.
• Add pagination to list_* tools: add optional limit (default 20, max 100) and offset parameters. Return {items: [...], total: int, offset: int, limit: int} so LLM can fetch subsequent pages.