Telegram AI Agent with LangGraph & Supabase

Interest in AI agents and conversational interfaces has surged recently, but most implementations remain confined to web apps or proprietary platforms. We wanted to explore a more open, user-friendly approach — one that runs inside a familiar, everyday tool. Telegram stood out as the ideal environment: it’s widely adopted, cross-platform, open source, and offers a powerful bot API. We built a Telegram-based agent system using Python and FastAPI. It integrates memory, tool calling, web search, speech processing, scheduled tasks, and user permissions — all controllable through natural language, directly within chat.

We used LangGraph to manage the agent’s flow, OpenRouter for LLM model access, AssemblyAI for speech-to-text, and Tavily for web search. For memory and state, we used Supabase. All of this runs in a small Docker setup. We’ve separated the Telegram bot handling, background processing, and scraping tools into individual services so the system remains modular and easy to adapt. If you want to deploy it, it’s straightforward to run locally or host with something like Coolify.

Tech Stack

Telegram Bot API

Provides a robust, free platform with excellent documentation and virtually limitless possibilities for bot development.

FastAPI

Modern, fast web framework for building APIs with Python. Provides automatic interactive API documentation, data validation, and excellent performance for handling webhooks and backend services.

LangGraph

Significantly simplifies developing comprehensive agentic systems. It provides sophisticated workflow orchestration that moves beyond simple chat completions to stateful, multi-step interactions. For production deployments, LangGraph Cloud offers enterprise-grade capabilities with enhanced scaling and monitoring.

OpenRouter

Fantastic for AI development - it provides unified API access to dozens of state-of-the-art language models from different providers. This makes it easy to switch between models, optimize costs, and implement fallback systems without managing multiple API keys.

Supabase

Offers PostgreSQL with real-time capabilities and excellent developer experience. We extended LangGraph’s checkpointing system to work with Supabase, creating a hybrid memory architecture where each user maintains one persistent thread with optimized short-term memory.

AssemblyAI

Provides advanced speech-to-text capabilities.

Tavily

Delivers reliable web search and research capabilities that integrate seamlessly with AI agents.

System Architecture

The system demonstrates several key integrations:

• Memory Management: Extended LangGraph checkpointing with Supabase for persistent, per-user conversation threads
• Multi-Modal Processing: Speech-to-text with intelligent summarization for voice messages
• Tool Integration: Web search, market data analysis, file processing, and task scheduling
• User Management: Natural language admin controls - just ask the bot to add or remove users
• Model Flexibility: Dynamic switching between different LLMs based on task complexity through OpenRouter

Getting Started

To try it locally, you’ll need to follow these steps:

Open Source

The complete codebase is available on GitHub under the MIT license - Check out the GitHub repository