Server has 4 tools with Spanish descriptions and basic schemas. Naming follows verb_noun pattern (auth_*, operate). Descriptions are present but lack LLM-optimized structure, they read as implementation notes rather than agent guidance. Parameters have types but descriptions are sparse or missing. No output schemas documented. Error handling delegates to domain layer via toolText() wrapper, which returns Spanish error messages but provides no recovery guidance. Tool composition is reasonable (auth_login, auth_status, auth_logout, operate are distinct concerns), but descriptions don't explain WHEN to call each or what prerequisites exist.

• Rewrite descriptions as agent guidance: 'auth_login: Opens secure browser window for SII tax ID entry. Call this first to establish session. Idempotent if session already active. Set www2=true to also authenticate for Portal MIPYME (required for carpeta_* tools).'
• Document output schemas for each tool. E.g., auth_status returns {authenticated: boolean, rut?: string, expiresAt?: string, www2: {authenticated: boolean, expiresAt?: string}}. Include in tool description or separate schema field.
• Add recovery guidance to error messages. Wrap domain errors with context: 'Session expired. Call auth_login to re-authenticate and retry.' instead of raw Spanish message.
• Expand parameter descriptions with constraints and examples. E.g., 'rut: Chilean tax ID (format: XX.XXX.XXX-K). Call operate list=true to see available RUTs.' instead of bare 'RUT of the entity to operate as'.
• Add tool prerequisites and sequencing hints. E.g., auth_login description: 'Prerequisites: none. Call before any other tool. Returns when user completes browser login flow.'