This server has significant definition quality gaps. All 14 tools have schemas with proper types and required fields, which is positive. However, descriptions are uniformly poor: most are 10-30 characters (e.g., 'Pobiera liste leadow z CRM' = 26 chars), well below the 50-200 char baseline for LLM-optimized descriptions. Polish-language descriptions provide no context on WHEN to use each tool, what it returns, or dependencies. Parameter descriptions are entirely absent, no tool documents what its parameters do. For example, 'create_lead' accepts 'pipeline_type' enum but never explains what cold_lead vs contact means. Output schemas are not documented anywhere. Error handling is absent: no guidance on retryability, user-fixable vs fatal errors, or recovery paths. The server lacks idempotency hints, permission declarations, and audit trail documentation. While schemas are syntactically valid, the lack of parameter and output documentation, combined with minimal tool descriptions, makes this server difficult for LLMs to use correctly.

• Expand all tool descriptions to 80-150 characters. Example: 'create_lead' should be 'Creates a new lead in the CRM system. Non-destructive operation. Requires company_name; other fields optional. Returns lead_id and confirmation.' This tells LLM what, when, and what it returns.
• Add descriptions to every parameter. For 'pipeline_type', write: 'Type of lead pipeline: cold_lead for new prospects, contact for existing relationships.' For 'data' in update_lead, specify: 'Object with fields to update (company_name, email, phone, status, etc.). Only provided fields are modified.'
• Document output schemas for all tools. Example for list_leads: 'Returns {leads: [{lead_id, company_name, email, status, created_at}], total: number, page: number}'. This enables LLMs to chain calls and extract needed fields.
• Add error handling guidance. Example: 'If lead_id not found, returns 404 with message "Lead not found. Try search_leads() with company name." This is not retryable, verify the ID and search again.'
• Add tool annotations in MCP schema. Mark create_lead, update_lead, add_activity, create_task with destructiveHint=true. Mark all list_* and search_* with readOnlyHint=true. This helps LLMs reason about side effects.