MCP for working with Telegram Bot API. Provides functions for sending and receiving messages through Telegram bot.
The server defines 5 tools with explicit FastMCP registration, all with descriptions and input schemas. Naming is action-verb based and mostly clear (send*, get*, delete*). However, descriptions are minimal (10-50 chars for most), parameter descriptions are present but terse, output schemas are not documented, and error handling is basic. The schema quality is moderate: all parameters are typed, but descriptions lack actionable detail about constraints, format, or prerequisites. Missing are output schema documentation, pagination/limit enforcement, and detailed error recovery guidance. No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels (READ_ONLY, WRITE, DESTRUCTIVE).
Deletes a message from the chat.
Gets information about the bot.
Gets updates (messages, events) from Telegram Bot API.
Sends a text message to the specified chat.
Sends a photo to the specified chat.
Output schemas not documented. All 5 tools return Dict[str, Any] but neither the response structure nor the fields are described. LLMs cannot plan downstream calls or extract required IDs without this information.
Tool descriptions are extremely brief (10-50 chars). Descriptions lack WHEN to use (vs similar tools), prerequisites, or state-modifying consequences.
No tool annotations despite declared risk levels. deleteMessage is DESTRUCTIVE; sendMessage and sendPhoto are WRITE. Per spec, destructiveHint=true and readOnlyHint=true are missing. This prevents the LLM from reasoning about irreversibility and safety.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 47 | <=2025-11-25 | v2 |
| 2026-03-09 | D | 50 | - | v1 |
Parameter descriptions are terse and lack constraint details. 'Chat ID where to send the message (string)' tells the LLM it is a string but not format (is it numeric? alphanumeric?). 'Message text to send' lacks length limits or markup hints. No enums for API constraints.
Error handling is generic. All tools return {success: bool, error: str} with no classification (retryable, user-fixable, fatal). Errors like 'TelegramError' are logged but not translated to actionable recovery hints. No guidance on what the LLM should do next.
getUpdates returns 'updates: [update.to_dict() for update in updates]' with no limit enforcement and no documented pagination. Unbounded results blow context windows.
TELEGRAM_BOT_TOKEN is environment-injected (good), but tool composition forces the agent to call getMe() before sendMessage(), LLMs cannot verify bot identity in one call. No batch operations for multi-chat messaging.