nanobot: A Lightweight Personal AI Agent You Host Yourself (GitHub, Scanned)

nanobot is a personal AI agent you run yourself, built by the University of Hong Kong's data intelligence lab (the team behind LightRAG) as a deliberately small, readable alternative to larger agent platforms. Install it, open the web UI, add a model, and you have an assistant that can work with files in folders you choose, search and read the web, run shell commands, generate images, delegate to sub-agents and remember your preferences across conversations.

It is built to stay running. A local gateway serves the browser UI and a terminal UI, and the same agent can answer in Telegram, Discord, Slack, WeChat, Feishu, Microsoft Teams, Mattermost, Linear or email. Scheduled automations and long-running goals let it check things and report back on a timer, and it exposes a Python SDK and an OpenAI-compatible API for your own integrations. It works with OpenAI-compatible providers, local model servers and fallbacks between them, connects to MCP servers, and has a Restricted mode that limits file access to a folder.

It is MIT-licensed, with about 49,300 stars and nearly 500 contributors, at v0.3.5 when scanned, and installs from PyPI as nanobot-ai.

• Repository: github.com/HKUDS/nanobot
• Licence: MIT (MIT License)
• Language: Python. Stars: 49.3K. Forks: 8,661. Last push: Oct 22, 2026.
• Scan: safe, Oct 22, 2026, commit 421bd6b

Who it is for

People who want an always-on assistant in their own chat apps without handing their files to a hosted service, tinkerers who found bigger agent frameworks hard to read, and developers who want a small agent core to extend.

Getting started

1. Install the stable release with uv (or python -m pip install nanobot-ai in a virtual environment)

uv tool install nanobot-ai

2. Open the browser UI and add a model under Settings, Models

nanobot webui

3. Or set it up from the terminal instead

nanobot onboard --wizard

The project also offers a one-line installer (curl | sh on macOS and Linux, irm | iex on Windows) that accepts --dry-run to preview what it will do. Start file work in Restricted mode on a practice folder, and check each chat platform's allowlist settings before connecting the bot to shared channels.

Safety scan

We cloned HKUDS/nanobot at commit 421bd6b on Oct 22, 2026 and ran the checks described on the GitHub Tools page: credential patterns, decode-and-execute code, install-time scripts, committed binaries, risky CI workflows, every host the code talks to, known vulnerabilities in pinned dependencies, and project hygiene. A person read every hit. This is what we found.

• No real secrets across 2,154 files and about 615,000 lines of Python and TypeScript; the four hits are fake Slack and AWS tokens in web UI tests.
• The pattern hits are explained. The exfiltration-host hits are the Telegram channel validating your bot token against api.telegram.org, its documented purpose. The PowerShell -EncodedCommand call opens the web UI in your browser from Windows or WSL. The rest are tests, including ones that check the agent blocks a piped shell download.
• The installers are long but tame: about 650 lines each for install.sh and install.ps1, no sudo, no shell profile edits, a dedicated environment under ~/.nanobot, and a --dry-run flag to preview the plan.
• 44 known advisories (4 critical), all in webui/package-lock.json and in test tooling (Vitest, happy-dom, tinypool) rather than in what the browser loads. The Python package has no lockfile in the repository.
• 5 workflows, none using pull_request_target; 3 of 13 third-party actions pinned to commits. Security policy, licence and contributing guide present; no Dependabot or CodeQL.

What the scanner counted

The raw findings

Every hit the scanner wrote out, with a link to the exact line at the scanned commit. Secrets candidates are redacted.

By the numbers

Lines by language: Python 356.8K, TypeScript 176.1K, JSON 54.8K, Markdown 19.8K, CSS 2,279, YAML 1,439.

Questions

Is nanobot free?

Yes. nanobot is MIT-licensed and free to self-host. You pay for the model provider you connect, or nothing with a local model server; the project itself sells nothing.

How is nanobot different from OpenClaw or Hermes Agent?

All three are self-hosted personal agents that live in chat apps. nanobot aims to be the small, readable one: a compact Python core with the web UI, memory, automations and MCP built in, which makes it easier to understand and modify. The larger projects have more integrations and bigger communities.

Can nanobot use a local model?

Yes. Point it at Ollama, LM Studio, vLLM or any OpenAI-compatible server in the model settings. Agent work with tools and memory needs a capable model with tool calling, so very small local models will be limited.

This post is part of GitHub Tools, where every repository is cloned and scanned before it is written up. The scan is a snapshot of one commit on one day; the repository has moved on since, so check it before you install.