nanobot: The Personal AI Agent You Can Actually Own

You want an AI assistant. Not Claude on a web page — a real assistant that’s yours, runs where you want, talks to you on Telegram when you’re away from the computer, and remembers what you talked about last week. That’s what nanobot is. And it took me under 5 minutes to set up.

I’ve tested three flavours of AI agent this month. Goose (48K★, Linux Foundation) is a terminal-first CLI agent — great if you live in the shell. Odysseus (56K★) gives you a full Docker workspace with ChromaDB, SearXNG, and email. But nanobot (44K★, HKUDS) sits in a spot neither quite fills: a chat-native personal agent that reaches you wherever you already talk.

What nanobot Actually Does

Here’s the quick version. You pip install nanobot-ai, run the setup wizard, and you’ve got an AI agent ready to go. No Docker compose, no frontend build, no config file rabbit hole.

The core is refreshingly small and readable — it’s Python, MIT-licensed, and the agent loop is about 500 lines. But the feature set punches way above its weight:

• WebUI ships inside the wheel — open http://localhost:8931 and you’re chatting
• MCP protocol support — plug in custom tools and data sources
• Model routing — set primary + fallback providers, it switches automatically
• Dream memory — two-stage memory that persists context across sessions
• Cron automation — natural-language scheduling, like “remind me every Monday”
• /goal system — sustained objectives that survive multi-turn conversations

But the killer feature? The channel list.

I connected Telegram first. Dropped my bot token into the config, restarted nanobot, and my agent was responding in my DMs within 30 seconds. No ngrok, no webhook setup, no reverse proxy — the app handles channel bridging internally.

Hands-On: The Install Flow

I ran this on a Ryzen 9 workstation running Ubuntu 24.04, Python 3.12.

pip install nanobot-ai
nanobot onboard --wizard

The wizard walks you through setting up your primary LLM provider (OpenAI, Anthropic, DeepSeek, local Ollama — you pick), gives model recommendations, and writes the config. Took maybe two minutes.

Then:

nanobot agent -m "What's the weather in Tokyo?"

First reply came back in about 4 seconds via GPT-4o. Clean, threaded, no setup fuss.

Then I opened the WebUI at localhost:8931 — it’s a proper chat interface, not a debug panel. Session history, model selector, tool outputs visible. It looks and feels like a product, not a side project.

How It Compares

The table above shows the channel breadth, but the real difference is where you interact with each tool.

nanobot is the one I’d recommend to a friend who says “I want an AI assistant I can talk to everywhere.” Goose is for the CLI power user who wants agentic code editing. Odysseus is for the stack builder who wants a full AI lab in a box.

What to Watch Out For

It’s not all roses. nanobot is barely 4 months old (first release Feb 2026) and the pace of change is relentless — daily commits, weekly releases, breaking config changes between minor versions. If you set it up today, the setup wizard might look different next month.

The Dream memory system is clever but still feels beta. Long-running sessions can get sluggish after 100+ messages if you don’t enable auto-compact. And the multi-language documentation is extensive but uneven — the English docs are solid, but some channel-specific guides lag behind the code.

One thing that surprised me: there’s no official Docker image in the main repo. The team recommends Docker deployment via a community wiki page, which works fine but feels like an afterthought for a tool that clearly benefits from being always-on.

The Bottom Line

nanobot is the closest thing to “personal AI assistant, out of the box” I’ve seen in open source. 44K stars in 4 months isn’t hype — it’s people who actually want this and found it. If you’ve been waiting for an AI assistant that’s yours, not rented, this is a top starting point today. And if you want it running 24/7, a $6/month VPS is all you need to keep it alive (affiliate link).