Server has 10 tools with complete input schemas and consistent naming (verb_noun pattern: telegram_send, telegram_get_*). All tools include descriptions (avg ~80 chars) and parameter descriptions. However, output schemas are entirely undocumented, no response structure is defined anywhere in the code. Error handling is minimal: no recovery guidance, no categorization of retryable vs fatal errors, no actionable error messages. The 'companion' parameter is well-designed with enum constraints, but most other parameters lack format/range constraints. No tool annotations (readOnlyHint/destructiveHint) despite clear READ vs WRITE semantics. Missing: pagination support for telegram_get_updates (could return hundreds of updates), output field documentation, and error recovery patterns.

• Document output schemas for all tools. For send_* tools, specify: { success: boolean, message_id: number, timestamp: string, error?: string }. For get_* tools, specify the exact fields returned (e.g., telegram_get_me returns { id, is_bot, first_name, username, ... }).
• Add error handling and recovery guidance. Wrap telegramRequest() calls in try-catch; categorize errors (invalid_chat_id → 'Chat not found. Verify chat_id is correct.', rate_limited → 'Rate limited. Retry in 30s.', network_error → 'Network error. Retry.') and return actionable messages.
• Add pagination to telegram_get_updates: accept limit (1-100, default 10) and offset parameters; document max 100 results per call; return total_count and next_offset in response.
• Add tool annotations: mark all send_* tools with destructiveHint=true; mark all get_* tools with readOnlyHint=true. This signals to agents which tools have side effects.
• Enhance parameter descriptions with format/range constraints: chat_id → 'Telegram chat ID (numeric or negative for groups)', message → 'Text message (max 4096 chars)', limit → 'Number of updates (1-100, default 10)'.