The server defines 3 tools with explicit JSON schemas and basic descriptions. All tools have input schemas visible in the code with typed properties and required field declarations. Naming follows verb_noun convention (create_task, create_webhook, delete_webhook). However, descriptions are minimal (34-67 chars), parameter descriptions are sparse or missing, output schemas are undocumented, and error handling lacks recovery guidance. Schemas are present but lack depth in field constraints and validation rules. No tool annotations (destructiveHint, readOnlyHint, idempotentHint) despite delete_webhook being destructive.

• Expand tool descriptions to 80-150 characters each, explaining WHAT the tool does, WHEN to use it (vs similar tools), and what dependencies or prerequisites exist. E.g.: 'Create a new AI task in Manus for processing text, images, or documents. Use this to offload long-running or complex analysis tasks. Returns task_id and status tracking URL.'
• Add full parameter descriptions for all inputs. For 'mode', explain: "Execution mode: 'speed' prioritizes fast turnaround (suitable for simple queries); 'quality' prioritizes accuracy with longer processing (suitable for complex reasoning tasks)." For 'events' in webhooks, document: "Event types to subscribe to (e.g., 'task.completed', 'task.failed', 'task.progress'). Omit to subscribe to all events."
• Document the response schema for each tool. E.g., create_task should return: {task_id: string, task_title: string, task_url: string, shareable_link?: string, created_at: string}. Add a 'Returns' section to each tool description or create a separate schema definition file.
• Add tool annotations: mark delete_webhook with destructiveHint=true. Consider marking create_task and create_webhook with idempotentHint=false if they lack deduplication (LLM should not retry without user confirmation).
• Implement structured error responses. Replace try-catch errors with categorized responses: {error_code: 'INVALID_MODE', message: 'Invalid mode. Must be one of: speed, quality.', suggestion: 'Retry with a valid mode.'} This enables LLM self-correction.
• Add a confirmation pattern for delete_webhook. Return {status: 'pending_confirmation', webhook_id: '<id>', message: 'Confirm deletion of webhook <url>?'} or implement a dry_run parameter.