llm_tg_bot

Local Telegram AI-agent on your own model. No cloud.

Telegram: @cmonjustdoitbot

Source: https://github.com/Fakeonomics/llm_tg_bot

Features

• Agent with tools — web search, page fetch, files, sandbox
• Fast-path: simple questions answered in 1 generation
• Real rich formatting (Bot API 10.1+) — tables, headings, lists, code, quotes, spoilers
• Private chats, groups (by mention), guest mode, inline
• Model vision (mmproj) + OCR fallback
• Any OpenAI-compatible API (local / OpenRouter / Together / NVIDIA…), models auto-fetched
• Second API as fallback (server-side env only)
• Fair per-user queue + anti-spam cooldowns, live progress, cancel by stop-word
• Access for everyone or admin-only, switchable at runtime

Commands

Command	Description
/reset	Clear chat context immediately
/ctx	Context usage
/start	Help with all commands
/admin open|close|status	Access for all / admin-only / status (admin)
/models	List models from the active API (admin)
/setapi <url> [key]	Switch backend (admin)
/setmodel <id>	Set model manually (admin)
/setcd [private group]	Anti-spam cooldowns (admin)

Terminal:

Command	Description
tgbot [start|stop|restart|status|logs]	Manage the bot

Install

git clone https://github.com/Fakeonomics/llm_tg_bot && cd llm_tg_bot
./scripts/install.sh

install.sh creates venvs, installs pinned deps, asks for BOT_TOKEN and admin ID, auto-detects the model port and remembers it. Then:

tgbot           # start (background)
tgbot status    # check
tgbot logs      # live log

Requirements: python3, a model on an OpenAI-compatible API (LM Studio / llama-server / hosted). Optional OCR: tesseract.

Usage

• Private: just write. Files and photos accepted.
• Groups: mention @bot or reply to its message.
• Guest mode: tag the bot even where it is not a member (needs Guest Mode ON in BotFather).
• Inline: @bot query in any chat.
• Stop: send stop/cancel or tap the Stop button under the working message.
• Research questions may take up to 5 tool steps; simple ones answer in 1 generation.

Configuration

.env files (never committed, see configs/*.example):

Key	Where	Meaning
BOT_TOKEN	yymbot/.env, base/.env	Token from @BotFather
ADMIN_LIST	yymbot/.env	Admin Telegram ID
whitelist	yymbot/.env	Empty = open to all
LLM_BASE_URL / LLM_API_KEY / LLM_MODEL	env	Active backend (set via /setapi)
EXTRA_BASE_URL / EXTRA_API_KEY / EXTRA_MODEL	env, server only	Second API fallback

Layout

bin/tgbot                    # launcher: start/stop/restart/status/logs
scripts/install.sh           # one-command setup
configs/*.example            # env templates
deployment/tgbot.service     # systemd unit example
yymbot/bot.py                # engine: private/groups/guests/inline
yymbot/smo_bridge.py         # bridge: queue, cooldowns, files, vision, delivery
base/src/services/smo_agent.py     # agent (CodeAgent + native LFM loop)
base/src/services/agent_loop.py    # delivery + privacy scrub
base/src/services/llm_config.py    # backend registry (any OpenAI-compatible API)
base/src/services/richdraft.py     # rich: markdown to JSON blocks
base/src/services/status.py        # live status, quiet net-error logging

Notes

• Slow answer? Check send_final via= and private lag/done in logs: network vs model.
• Old Telegram clients may render rich tables as text — that's the client, not the bot.