HKUDS/nanobot

Nanobot is an ultra-lightweight personal AI agent framework from HKUDS. 21,000+ stars on GitHub as of April 2026. MIT license. Python. One small core loop.That is the entire pitch. And it is working.What It Actually IsNanobot is a single-loop agent runtime. You define channels (conversations), you plug in memory (SQLite by default), and you wire in MCP servers for tools. The core loop handles message routing, context assembly, tool calls, and memory writes. That is it.No supervisor hierarchies. No team orchestration. No state machines. No DAG builders. No YAML config. Just Python.The minimalism is the feature. Every agent framework that grew into an enterprise stack — LangGraph, CrewAI, AutoGen — started simple and accumulated abstractions. Nanobot is betting that the simple version is what most personal-agent use cases actually need.What You Get Out of the BoxChat channels. First-class conversation scoping so one Nanobot instance can run 50 parallel conversations without context bleed.Persistent memory. SQLite-backed long-term memory with automatic summarization on overflow. No vector store required for the default path.Native MCP support. Drop any MCP server into the config and Nanobot treats its tools as first-class agent actions. Works with every MCP server in the ecosystem — file system, GitHub, Slack, Notion, browser automation.Model-agnostic. Point it at OpenAI, Anthropic, DeepSeek V4, Gemini, or any OpenAI-compatible local endpoint. Swap in a config line.Readable source. The core loop is under 2,000 lines. You can audit the entire codebase in an afternoon.Who Should Use ItUse Nanobot if you want a personal agent that runs on your laptop, remembers you across sessions, talks to the tools you use daily via MCP, and does not require a 400-page framework tutorial. It is the right answer for solo devs, hobby projects, personal automation, and teams prototyping agent ideas before committing to enterprise orchestration.Skip it if you need multi-agent supervisor hierarchies, complex state machines, or enterprise audit logs. Nanobot is deliberately not that. For those use cases, LangGraph or CrewAI still win.Install and First Runpip install nanobot-ai from nanobot import Nanobot bot = Nanobot( model="deepseek-v4-pro", memory="sqlite://./nanobot.db", mcp_servers=["filesystem", "github"] ) bot.chat("Summarize my open GitHub issues and write a plan.")That is a complete working personal agent with memory and tool access. Swap deepseek-v4-pro for any OpenAI-compatible model. Add or remove MCP servers in the list. Done.Why the Star Count Matters21,000 stars in under a year is the kind of signal that usually means one of two things: the framework genuinely works for the median developer, or it got a viral moment and will flame out. Nanobot is the former. The star curve is steady, the issue tracker is healthy, and the code is readable by anyone with a year of Python experience.In a 2026 AI ecosystem drowning in enterprise-first agent frameworks, Nanobot is the personal-first one that just works. Install it this weekend.Related ResourcesArticle: DeepSeek V4 just launched — the open-weight frontier model you should pair with Nanobot today.Tool: DeepSeek V4 — plug straight into Nanobot's model config for a frontier-grade personal agent at $3.48/M output.MCP server: Vercel Next.js DevTools MCP — one of many MCP servers Nanobot wires in via a single config line.Skill: Anthropic Cybersecurity Skills — 754 skills Nanobot can expose as tools to any agent loop.