Basavaraj_nick's Telegram-native Multi-Agent AI System

Problem

I was juggling six different tools just to get through a normal day — checking briefings here, deploying code there, running LLM queries somewhere else, generating images in another tab, and manually checking my calendar on top of it all. There was no single place I could just send a command from and have it actually get done. I wanted one always-on agent I could talk to from anywhere, with zero manual setup every time.

Results

This is my daily driver now — 14+ working commands spread across 8 sub-agents, running entirely on free-tier infra (Cloudflare Workers, KV, D1, and GitHub Actions). Zero paid hosting since launch, and it's still running strong.

Challenges

• Getting a 5-provider LLM fallback chain working (Mistral → Cerebras → Groq → OpenRouter → Gemini) that quietly degrades instead of throwing errors at me when one provider goes down.
• Making a stateless webhook pipeline on Cloudflare Workers actually handle long-running Python sub-agents, without blowing past the 50ms CPU limit.
• Keeping every API key and token out of the repo entirely, while still making them usable across Workers, GitHub Actions, and my local dev environment without constant friction.
• Building 5-minute rolling conversation memory in Cloudflare KV so the bot feels like it's actually following the conversation, without storage costs spiraling.

Architecture

At the core is a Python orchestrator running 8 specialized sub-agents — one each for briefings, deployments, LLM queries, image generation, memory, calendar, weather, and web search. Telegram is the only interface — no dashboard, no app, just commands typed into a chat. Cloudflare Workers handle the serverless webhook layer, Cloudflare KV stores short-term conversation memory, and GitHub Actions runs the daily scheduled tasks. The whole thing runs without a single dollar of paid hosting.