Reference
Nanobot: Two Different Projects Share the Name
“Nanobot” currently points at two unrelated open-source projects, and search results mix them freely. One is a compact Python agent framework inspired by OpenClaw. The other is an MCP host that turns MCP servers into agents. Picking the wrong one costs an afternoon, so here is the distinction plainly.
The two projects side by side
| HKUDS/nanobot | obot-platform/nanobot | |
|---|---|---|
| Origin | HKUDS lab, University of Hong Kong | Obot AI |
| What it is | Self-hosted personal agent framework | MCP host for building MCP agents |
| Written in | Python | Standalone host, deployable or embedded |
| Core idea | OpenClaw's capabilities in a tiny readable core | Wrap any MCP server into a conversational agent |
| Interfaces | WebUI, terminal, chat apps | Chat window with MCP-UI components |
| You want it when | You intend to read and modify the agent loop | You already have MCP servers and want a front end |
HKUDS/nanobot: the small readable agent
This one is explicitly inspired by OpenClaw. Its pitch is that you can get the core of a personal agent — tools, long-term memory, MCP integrations, model routing, multi-agent delegation, scheduled automation and an OpenAI-compatible API — in a codebase small enough to read in an afternoon.
That is a genuine strength for research and for learning how agents work, and it is the wrong lens for judging it against production software. A small core is small partly because it does less: the long tail of channel adapters, retries, session handling and operational hardening is exactly the code that makes a framework large. We wrote a fuller comparison in OpenClaw Launch vs Nanobot.
obot-platform/nanobot: the MCP host
Different problem entirely. You have MCP servers — passive collections of tools — and you want something a person can talk to. This host supplies the missing half: reasoning, a system prompt, tool orchestration, and MCP-UI support so a tool can render an actual interface component in the chat rather than returning text.
It handles both MCP transports: stdio for local servers and HTTP for remote ones. If you have been building MCP servers and testing them through a developer tool, this is the step that turns them into something you would hand to a colleague.
obot-platform/nanobot (the old nanobot-ai/ URL redirects there) and is in maintenance mode — not accepting new features or external contributions, with new development pointed at the MCP Go SDK instead. It still runs, and for evaluating an idea that is fine. For something you intend to operate for a year, a project that has stopped taking features is a different proposition, and worth knowing before rather than after.How to tell which one a page is about
- Look for the repository owner.
HKUDS/orobot-platform/(formerlynanobot-ai/) settles it immediately. - “Inspired by OpenClaw” or a percentage-of-code claim means the HKUDS framework.
- “MCP host”, “wrap an MCP server” or “MCP-UI” means the Obot AI project.
- Python-first framing with a WebUI points at HKUDS.
What carries over to any agent
Whichever you experiment with, the durable artifact is your MCP servers. MCP is a shared protocol, so a server you write while evaluating either nanobot registers unchanged with OpenClaw or Hermes Agent. Framework choices are reversible; the tools you build are the part worth keeping.
If you do not want to maintain a framework
Both projects are self-hosted, which means a server, updates, and being the person who notices when it stops at 3am. If what you want is an agent that stays online across Telegram, Discord, WhatsApp and other channels, OpenClaw Launch runs that for you and still lets you attach your own MCP servers and model keys.
Nanobot FAQ
Are there really two projects called nanobot?
Yes, and they are unrelated. HKUDS/nanobot is an ultra-lightweight self-hosted personal agent framework in Python. obot-platform/nanobot, from Obot AI, is an MCP host for turning MCP servers into agents. Same name, different problems, different repositories.
Which nanobot is the OpenClaw-inspired one?
HKUDS/nanobot, from a lab at the University of Hong Kong. It is explicitly inspired by OpenClaw and aims to deliver core agent capabilities in a fraction of the code — a research-friendly, readable core rather than a production replacement.
What does the MCP-host nanobot actually do?
It wraps an MCP server with reasoning, a system prompt and tool orchestration, so a passive set of tools becomes an agent you can talk to. It supports local stdio and remote HTTP MCP servers and can render MCP-UI components in the chat window.
Can I use either one with OpenClaw Launch?
Not as a drop-in. Both are separate runtimes rather than plugins. What does transfer is any MCP server you built for either — MCP is the shared protocol, so the same server registers with OpenClaw or Hermes Agent unchanged.
Which should I pick?
If you want to read and modify an agent's core, take the HKUDS framework. If you already have MCP servers and want a conversational front end over them, take the Obot AI host. If you want an agent running on your messaging channels without maintaining either, that is a managed service instead.