AI Secretary

Telegram AI Secretary connects to a Telegram account using the Telethon library, intercepts incoming messages before they reach the native notification layer, and applies a two-stage filter: first a rule set defined in notification_rules.py, then an LLM call for messages where context is ambiguous. Messages that clear both stages are forwarded to your phone via ntfy. The entire stack runs on your own hardware — a Raspberry Pi or any always-on Linux machine with Python is sufficient, and a systemd unit file is included for daemonization.

The differentiating design choice here is the LLM-in-the-middle evaluation step. Rather than keyword lists or sender whitelists alone, the tool passes message content to a language model to judge urgency — catching phrasing like ‘I’m outside’ or ‘are you free in an hour’ that rigid rules miss. The vendor describes this as filtering for ‘calls, same-day plans, or someone waiting at the door,’ which maps to exactly the cases keyword filters fail on.

This tool fits a narrow but real scenario: a technical solo user or small household who keeps Telegram muted by default and wants an LLM to act as a triage layer. It breaks when the filtering logic needs to account for conversation history — each message is evaluated in isolation, so a follow-up message like ‘seriously, I’m waiting’ with no prior context in the current session may not pass the filter. Teams managing multiple users, multiple platforms, or who need a configuration UI will hit the ceiling of a single-file Python project quickly and move to a hosted notification management service instead.

Dependencies are declared in requirements.txt. The .env.example file covers Telegram API credentials, LLM endpoint configuration, and ntfy target. No database is required; notification_state.py handles lightweight in-process state. MIT licensed, with no paid tiers and no external service dependency beyond whatever LLM endpoint you configure.