What is nanobot?
nanobot, from the HKUDS lab at the University of Hong Kong, is a small, readable, self-hosted agent framework in Python. One gateway owns execution, sessions, tools, memory and security policy, and the WebUI, native terminal UI and chat channels connect to it. The channels include Telegram, Discord, Slack, WeChat, Feishu, Email, Mattermost, Teams and Linear. It works with files (including PDF, Word, Excel and PowerPoint), shell, web search, MCP, cron schedules, image generation and delegation to sub-agents. Long-term memory is consolidated by a process it calls Dream.
Created in February 2026, the repo has about 49,700 GitHub stars and is MIT-licensed. Version 0.4.0 shipped on 28 October 2026, and the nanobot-ai PyPI package had about 19,900 downloads in the latest seven days. It needs Python 3.11+ and a model provider or local model server; it supplies no model account of its own. It also exposes a Python SDK and an OpenAI-compatible API.
Is nanobot worth it?
nanobot's case is size. It does much of what OpenClaw does, including persistent memory, scheduled tasks, chat-app reach and MCP, from a Python core small enough to read. For a developer who wants an always-on agent they can audit or fork, that matters more than OpenClaw's longer channel list. The Python SDK and OpenAI-compatible API also make it easy to call from existing code, which OpenClaw's TypeScript gateway does less directly.
The defaults are sensible. The first-run file task uses workspace-only permission, temporary chats stay out of history and memory, and the README warns that WebUI access means administering the agent instance. It also notes that approving a private sender on a chat platform does not always restrict group access. Those are the right warnings. They are warnings, though: the operator still has to configure pairing and group policy channel by channel.
The security history is short but not clean. Five GitHub advisories have been published, two critical and three high, all between March and May 2026. They include an authorisation bypass through injected sender IDs, cross-site WebSocket hijacking in the WhatsApp bridge, and a path traversal that allowed arbitrary file writes from WhatsApp document names. All were in the code that faces chat apps, which is the part a personal agent cannot avoid exposing.
The project moves very fast. The README's news log shows a release note almost every day in late October 2026, and there are about 870 open issues and pull requests. The source and stable tracks are documented separately for that reason. Teams that want to stand still should pin a PyPI release and read the updating guide before each upgrade.
Where it earns its place
A Python developer with a home server
Run a private assistant on Telegram that remembers preferences and runs nightly jobs
uv tool install nanobot-ai, connect Telegram, and add cron automations; memory lives in plain files such as USER.md and memory/MEMORY.md that can be read and edited. Free software; a local model server or API key supplies the model.
A two-person consultancy on Slack
Have an agent summarise client reports and post a Monday digest
The built-in file tools read the .docx, .xlsx and .pptx files, and a scheduled automation posts to Slack. Lock the workspace to the client folder and check the channel's group-access policy before inviting the bot.
A backend engineer adding an agent to an internal tool
Call a persistent agent with tools and memory from an existing service
Use the Python SDK in-process or point an OpenAI client at nanobot's OpenAI-compatible API. Costs are model tokens; pin a PyPI version, since the source track changes almost daily.
Before you commit
- Setup
- Run the install script (curl or PowerShell), which creates ~/.nanobot/venv and opens the WebUI for model setup; or uv tool install nanobot-ai or pip install nanobot-ai, then nanobot webui. Python 3.11+. A first reply takes minutes once a provider key is entered.
- Integrations
- Telegram, Discord, Slack, WeChat, Feishu, Email, Mattermost, Teams and Linear; MCP servers and presets; OpenAI-compatible and local model providers with fallbacks; Langfuse tracing; Python SDK and OpenAI-compatible API.
- Data and privacy
- Sessions, memory and files stay on your machine; prompts go to the configured model provider. Five GitHub security advisories (two critical, three high) were published between March and May 2026. MIT licence.
- Cost at scale
- No paid edition. Cost is model usage plus a small always-on machine or VM for the gateway.
- Limits
- Fast-moving source track and about 870 open issues and pull requests. WebUI access grants full administration. Per-platform group policies must be configured separately.
What’s good
- Small Python core that is practical to read and fork
- Python SDK and OpenAI-compatible API
- Workspace-only permissions in the default first-run flow
- WebUI, native TUI and 9+ chat channels
What’s not
- Two critical and three high advisories in its first months
- Near-daily changes; pin a release for stability
- Group-access rules must be set per channel
- Fewer channels and apps than OpenClaw
Best for
Python developers who want a self-hosted personal or team agent they can read, extend and call from their own code.
nanobot alternatives
nanobot: common questions
- Is nanobot worth it?
- nanobot is the personal agent to choose when you want to read the code you are trusting, smaller and easier to follow than OpenClaw, with a younger and spikier security record. We rate it 4.0 out of 5.
- How much does nanobot cost?
- Free and open source (MIT). You pay your model provider, or run a local model.
- What are the downsides of nanobot?
- Two critical and three high advisories in its first months; near-daily changes; pin a release for stability; group-access rules must be set per channel; fewer channels and apps than OpenClaw.
- Who is nanobot best for?
- Python developers who want a self-hosted personal or team agent they can read, extend and call from their own code.
- How much does nanobot cost for a team or in production?
- No paid edition. Cost is model usage plus a small always-on machine or VM for the gateway.
- Is nanobot safe for company data?
- Sessions, memory and files stay on your machine; prompts go to the configured model provider. Five GitHub security advisories (two critical, three high) were published between March and May 2026. MIT licence.
- What are the best nanobot alternatives?
- The closest alternatives we have reviewed are OpenClaw, Hermes Agent, OpenBot and goose.
Answers checked Oct 31, 2026.
Make nanobot? Show this rating on your site
