This is a pydantic-ai integration library, not a standalone MCP server. Tool definitions are inferred from apcore module descriptors at runtime via _descriptor_to_tool_def(), not explicitly registered in source. Schemas are present but minimal. Descriptions vary widely: crm_read and send_invoice have adequate descriptions (20-40 chars), but lookup lacks any description, and recent_activity's description is meta-commentary ('Needs the agent run context...') rather than actionable guidance. No parameter descriptions visible in schema definitions. No output schemas documented. Error handling deferred to apcore pipeline with no recovery guidance surfaced to LLM. Tool composition is reasonable (single responsibility), but naming lacks consistency (crm_read vs send_invoice vs wire_transfer, mixed verb_noun patterns). No evidence of idempotency markers, confirmation gates for destructive ops, or rate limiting.

• Add explicit parameter descriptions to all tools. E.g., for crm_read: 'record_id: The unique identifier of the CRM record to retrieve (string, 1-50 chars, alphanumeric).'
• Rewrite recent_activity description from meta-commentary to actionable guidance: 'Retrieve recent activity log for a customer, including timestamps, action types, and actors. Requires customer_id.'
• Document output schemas for all tools. E.g., crm_read should specify: 'Returns {id, name, email, created_at, status}' so LLM knows what fields to expect.
• Add confirmation gates for destructive operations. Implement a two-step pattern: wire_transfer_preview (dry-run) → wire_transfer_confirm (execute with confirmation_token).
• Standardize tool naming to verb_noun pattern. Rename crm_read → read_crm_record, crm_delete → delete_crm_record for consistency with send_invoice and wire_transfer.