Critical structural deficiency: 14 of 15 tools use a generic 'action' parameter with minimal schema and vague descriptions. This violates the single-responsibility principle and forces LLMs to guess at valid action values. Tool names (holded_contacts, holded_sales, etc.) are domain nouns, not action verbs, they do not convey what happens when called. The holded_discover tool documents this anti-pattern explicitly: 'call it with no action to see the actions in its error message.' This is a discovery failure, not a feature. Only holded_request has a slightly more specific schema (method, path). Descriptions are generic ('action parameter for managing X') and under 50 chars, providing no guidance on which actions exist, when to use them, or what they return. No output schemas are documented. Error handling relies on runtime discovery via error messages rather than upfront schema clarity.

• Replace the 'action' parameter with explicit, verb-prefixed tools: create_contact, list_contacts, update_contact, delete_contact, search_contacts, etc. Each tool should have a single, clear responsibility and a specific schema.
• Rename tools to start with action verbs: get_contacts, list_sales_documents, create_purchase, search_payments, update_treasury_account, etc. This enables LLMs to infer intent from the name alone.
• Expand descriptions to 100-200 characters explaining WHAT the tool does, WHEN to use it, and WHAT it returns. Example: 'Create a new contact in Holded. Returns contact_id, name, email, and phone. Use this before creating sales documents for a new customer.'
• Document output schemas for every tool. Specify which fields are returned, their types, and which are required for downstream tool calls. Include pagination info (limit, offset, total_count) for list operations.
• Add enum constraints to parameters that accept a fixed set of values. For example, if 'action' can be 'list', 'create', 'update', 'delete', declare it as an enum in the schema. This eliminates runtime discovery and guides LLM selection.