Vizo AI
Back
Research analysis

HKUDS/nanobot

August 7, 20263 min read

HKUDS/nanobot

Imagine a personal AI that lives on your laptop, remembers every conversation, calls out to the tools you already use, and can even run a tiny team of specialized agents — all without leaving your own network.

That’s not a sci‑fi fantasy. It’s exactly what the HKUDS/nanobot project delivers: an ultra‑lightweight, open‑source, self‑hosted personal AI agent framework written in Python. It comes with a built‑in WebUI, tool‑registration hooks, memory persistence, an MCP (Model‑Control‑Protocol) layer, multi‑agent workflow support, automation scripts, and pluggable chat adapters. In short, it’s a ready‑made chassis for building any AI assistant you can imagine — from a grammar‑checking “Catbot” to a full‑blown research aide.

Why a self‑hosted framework matters

  1. Data stays yours – No API keys, no third‑party servers. Sensitive context lives on your machine.
  2. Zero vendor lock‑in – You own the code, the model, the upgrade path.
  3. Extensibility by design – Add a new tool by dropping a Python function into the MCP registry; the framework auto‑wires it to every agent that needs it.
  4. Lightweight footprint – A few megabytes of dependencies, no Docker daemon required, and it runs on any OS that supports Python 3.9+.

These first‑principles advantages are why researchers, indie developers, and even small teams are gravitating toward nanobot when they need a “personal AI” that feels truly personal.

The recent 4‑hour Agentic AI Engineering workshop

Last week I ran a full‑scale, hands‑on workshop titled “Agentic AI Engineering: MCP, CrewAI, OpenAI Agent SDK.” The session walked participants through:

  • Model‑Control‑Protocol (MCP) – a simple JSON‑over‑HTTP contract for exposing any Python function as a callable tool.
  • CrewAI – a lightweight orchestration layer that lets you declare a hierarchy of agents, assign them roles, and let them pass state via shared memory.
  • OpenAI Agent SDK – for those who still want to tap into GPT‑4‑level reasoning while keeping the surrounding workflow fully self‑hosted.

The workshop culminated in a live demo where each participant deployed a custom “Catbot” into nanobot’s WebUI. Catbot was entered in DEV’s Summer Bug Smash (a Sentry‑powered bug‑fixing competition) and solved a subtle grammar‑parsing edge case that had tripped up larger language models. Within nanobot, Catbot:

  1. Registered its check_grammar function as an MCP tool.
  2. Gained a persistent memory slot to store user‑specific style guides.
  3. Started a chat session through the WebUI, where it could ask clarifying questions, call external spell‑checking APIs, and even hand off to a “Proofreader” agent when confidence was low.

Seeing a single Python file become a fully interactive, memory‑aware, multi‑agent participant in seconds was a light‑bulb moment for everyone in the room.

Takeaway: Build, own, iterate

If you’re looking to move beyond “chat‑only” AI experiences and into real‑world automation, nanobot gives you the scaffolding to:

  • Create modular agents that can be swapped, versioned, and tested independently.
  • Compose workflows where one agent’s output becomes another’s input, enabling planning‑execution‑reflection loops.
  • Expose any Python library (from Selenium to Pandas) as a first‑class tool without writing wrapper code.

Because the entire stack is open source and community‑driven, you can fork it, add features, or contribute bug‑fixes back — turning a personal experiment into a collaborative platform.

Next steps

  • Grab the repo: git clone https://github.com/HKUDS/nanobot.git
  • Spin up the WebUI: python -m nanobot.ui – you’ll see a minimal chat window ready for your first tool.
  • Join the community: We have a Discord channel and a weekly “Agent‑Design” livestream where developers share patterns and troubleshoot together.
  • Reserve a seat for the next 4‑hour workshop (dates announced on the repo’s README). It’s a hands‑on deep dive into MCP, CrewAI, and the OpenAI Agent SDK — exactly the kind of training that turns a curious coder into an AI‑agent architect.

The future of AI isn’t just about bigger models; it’s about smarter orchestration and ownership. With HKUDS/nanobot you get both, in a package that fits on a laptop and scales to a team. Let’s build the next generation of personal agents — together.

If you’re excited about self‑hosted AI agents, drop a comment below or ping me directly. I’d love to hear what you’d build on top of nanobot.