Server has 7 tools with mixed quality. Most tools have descriptions (good), but parameter schemas are incomplete or missing type information. voice_command and text_command lack input validation guidance. add_expense has complex split_map parameter with no type constraints or validation rules. delete_expense is destructive but lacks confirmation/dry-run pattern. No tool annotations (readOnlyHint/destructiveHint) present. Error messages are generic ('error: {e}') rather than actionable recovery guidance. Output schemas are undocumented, LLMs cannot predict response structure for chaining.

• Add tool annotations: mark delete_expense with destructiveHint=true, list_friends with readOnlyHint=true, add_expense with destructiveHint=true (modifies state). Enables agent safety checks.
• Document output schemas for all tools. E.g., add_expense returns {success: bool, expense_id: string, message: string}. list_friends returns {friends: [{name: string, id: string}]}. Enables LLM response parsing.
• Replace generic error messages with actionable recovery guidance. E.g., 'Configuration failed: invalid API key format. Expected 32-char hex string. Call login_with_token() if you have an OAuth token instead.'
• Add validation rules to split_map parameter description: 'Dict with keys=names (or "me"), values=amount (e.g. "10.50") or percentage (e.g. "50%"). All values must sum to 100% or total amount.'
• Move credentials (consumer_key, consumer_secret, api_key, access_token) to server-side environment variables. Expose a single configure_splitwise(config_name: str) tool that loads from env. Prevents credential leakage in logs.