This is a personal AI assistant that communicates with the user via Telegram. It’s a personal project that I decided to make public, so I’m forced to keep sensitive information (such as keys) outside of the source code. This project was started long before agents like OpenClaw existed, and it contains a mix of many then–state-of-the-art approaches to building an agent, though it is not always kept up to date. The project also includes many personal tools that help me with day-to-day organization.
The assistant runs in an agentic loop and is decoupled from the Telegram message-handling layer, so it can be invoked from anywhere in the codebase (not only from an incoming chat message). The agent communicates by calling send_telegram_message, and it ends the loop by calling submit_solution once it is done.
-
Slash commands registered in Telegram (appear in the
/command picker):/bounty- Toggle Torn bounty monitor on/off./clear_thread- Clear the AI conversation history (and memory graph, if enabled)./company- Open Torn company settings UI (currently training preferences)./daily_checkin- Manually trigger the daily habit check-in./email_here- Set this chat as the destination for email notifications./live_message- Create a message that updates every second (timer demo)./model- Switch the OpenRouter model used by the main agent./next- Show your next timetable lesson today./now- Show your current timetable lesson./ping- Reply withpong./pong- Reply withping./q- Save an encrypted private note (message is deleted after saving)./racing- Show Torn racing skill stats + predictions (with a graph when possible)./set_torn_api_key <api_key>- Set Torn API key for this bot instance./set_wolframalpha_app_id <app_id>- Set WolframAlpha App ID (stored in DB)./settings- Open notification/settings toggles UI./stacking- Toggle Torn stacking mode on/off./stock- Send a Torn stock report./target- Pick a suitable Torn target (targets list / hospital timing helper)./time_table- Add/manage timetable entries (interactive UI)./train- Send Torn training status./unwatch- Stop the active/watchfile watcher in this chat./watch [path]- Live-watch a file understorage/and auto-update the message on changes.
-
Utility slash commands (supported but not shown in the picker):
/cancel- Cancel an in-progress interactive flow (used by/watch,/model,/settings,/company,/time_table).
-
AI tools (functions the main agent can call):
seconds_until- Seconds remaining until a given datetime (%Y-%m-%d %H:%M:%S).convert_to_seconds- Convert days/hours/minutes/seconds into total seconds.create_reminder- Create a reminder that will message you later.cancel_reminder- Cancel reminders by ID.get_reminders- List active reminders.create_event- Create a Google Calendar event.send_telegram_message- Send a message to the primary Telegram chat (agent’s main output).read_file- Read a file fromstorage/with line numbers.write_file- Create/overwrite a file understorage/.str_replace- Replace a specific string in an existing file understorage/.list_directory- List directory contents understorage/.create_directory- Create a directory understorage/.delete- Delete a file/folder understorage/(recursive for directories).move- Move/rename a file/folder understorage/.search_files- Search filenames and contents understorage/.create_time_capsule- Schedule a “message to future self” delivery.create_habit- Create a new daily habit to track.list_habits- List active habits.remove_habit- Deactivate a habit.get_habit_stats- Get habit stats (streaks / averages / trends).generate_heatmap- Generate & send a habit heatmap image.submit_solution- Internal “done” signal for the agent loop.
-
Non-slash handlers (always on):
- Assistant chat: any non-command text/photo/voice message is routed to the main AI agent.
- File upload to memory: uploading a text-like file (txt/md/json/xml/js/yaml) adds its contents into memory.
This is all from time when function calling was like two weeks old, it has come long way since then and AI models are way smarter.
Looks like function output should be formatted in the style of json or at least provide context as a see better results when I do.
I can only guess this is because the AI doesn't remember what functions it called so when it calls for example get_current_time and gets only time like 10:20:34
it just remembers some numbers as result and doesn't know the context of them but if you return {"current_time":"10:16:19"} it will now know that what it got is current time.
Example of this: The AI really struggles to use get_current_timewith set reminder function to create reminders for specific times not in specific time.
After some more testing the problem could have been with the fact that I was returning time that had seconds as float making the AI not recognize it necessarily as time.
The AI has the ability to recognize errors and try again, but it seems this is very strict as it will only recognize outputs that specifically say that the function call failed. When your function returns result that is clearly incorrect it doesn't seem to recognize it as an error and will pretend the funtion did what it should have.
They provide nice speciality to them allowing them to perform specific tasks better than general assistant can. But the question is if they are needed or cheaper that just using GP-4, as this model seems to just be able to handle most task. Unfortunately because of the cost I can't test with the more advanced model.