Every side project starts the same way: "I'll just write a quick Telegram bot." Two hours later you're still debugging long polling, fighting with database schemas, and reinventing message chunking.
If you've ever built a Telegram bot connected to an LLM, you know the routine:
- Set up long polling or webhooks
- Handle message length limits (4096 chars)
- Create a SQLite schema for chat history
- Wire up an OpenAI-compatible client
- Add error handling so the bot doesn't die at 3 AM
None of this is hard. All of it is boring. And you rewrite it every single time.
The Solution: AgentKit
I packaged everything into a minimal Python boilerplate that gets you from git clone to working AI assistant in about 5 minutes.
GitHub: https://github.com/uDa4a100/agent-kit
The core design decisions:
1. Zero heavy dependencies
No aiogram, no python-telegram-bot, no openai SDK. Just urllib from the standard library and python-dotenv. Why? Because bot frameworks come and go, but the Telegram Bot API is just HTTP POST requests. Here's the entire API wrapper:
def _api(method, params):
url = f"https://api.telegram.org/bot{TELEGRAM_BOT_TOKEN}/{method}"
data = urllib.parse.urlencode(params).encode()
req = urllib.request.Request(url, data=data)
try:
with urllib.request.urlopen(req, timeout=40) as r:
return json.loads(r.read().decode("utf-8"))
except Exception as e:
LOG.error("Telegram API error: %s", e)
return None
That's it. Long polling is just calling getUpdates in a loop with a timeout parameter.
2. Any OpenAI-compatible backend
The config is one environment variable:
AI_BASE_URL=http://localhost:11434/v1 # Ollama
# or
AI_BASE_URL=https://openrouter.ai/api/v1 # OpenRouter free tier
Swap backends without touching code. Run a local model for privacy, or point it at a cloud API for power.
3. Message history in SQLite from day one
Every conversation is stored with timestamps:
CREATE TABLE IF NOT EXISTS messages (
id INTEGER PRIMARY KEY AUTOINCREMENT,
chat_id INTEGER NOT NULL,
role TEXT NOT NULL,
text TEXT NOT NULL,
ts TEXT NOT NULL
);
This matters more than people think. Once your messages are in SQLite, you get full-text search, analytics, and context injection for free later.
4. Non-blocking message handling
Each incoming message spawns its own thread:
threading.Thread(
target=_handle,
args=(chat_id, msg.get("text", "")),
daemon=True,
).start()
A slow LLM response doesn't block other users' messages.
Getting Started
git clone https://github.com/uDa4a100/agent-kit.git
cd agent-kit
cp .env.example .env
pip install -r requirements.txt
python main.py
Edit .env with your bot token from @botfather, pick an AI backend, done.
What's Next for This Project
The roadmap includes task scheduling (/add weather min:60 "check forecast"), FTS5-powered history search, and multi-user admin controls.
Links:
- Free version on GitHub: https://github.com/uDa4a100/agent-kit
- Full version (scheduler + FTS5 search + Docker): @agentkit_shop_bot on Telegram
What would you add to a personal AI assistant? Comments welcome.
