nanobot is an open-source ultra-lightweight personal AI agent from HKUDS, built in the spirit of OpenClaw, Claude Code, and Codex. With 42,000+ GitHub stars, 7,400+ forks, and 168+ releases, it has rapidly become one of the most actively developed personal agent runtimes of 2026. The project keeps the core agent loop intentionally small and readable while still shipping practical building blocks: chat channels, memory, MCP support, an OpenAI-compatible API, and ready-made deployment paths so a developer can go from local install to a long-running personal agent with minimal overhead. ## Why nanobot Matters Most agent frameworks have grown into heavy multi-layer stacks that are difficult to study, modify, or trust for personal use. nanobot takes the opposite approach. Its core is deliberately compact so researchers, hobbyists, and engineers can read the agent loop end-to-end, understand exactly what is happening on every turn, and modify it without fighting the framework. At the same time, it ships with the infrastructure most people actually need for a real personal agent: long-running sessions, persistent memory, scheduled reminders, multi-channel chat integration, and an API surface compatible with the broader OpenAI ecosystem. ## Multi-Channel Chat Integration nanobot natively connects to a wide range of communication platforms including Telegram, Discord, Slack, Feishu, Microsoft Teams, WeChat, WeCom, QQ, DingTalk, WhatsApp, Matrix, and email. A unified cross-channel session keeps context consistent across the platforms a user actually lives in, and channel plugins handle media, threads, typing indicators, reactions, and message splitting. This makes nanobot one of the most practical options for users who want a single agent reachable from every messaging app they already use. ## Memory, Reasoning, and MCP The agent ships with a redesigned token-based memory system, a two-stage Dream memory for skill discovery, and adaptive thinking support for providers like Anthropic and DeepSeek. Model Context Protocol (MCP) is a first-class citizen: multiple MCP servers can be configured at once, MCP resources and prompts are exposed as tools, and custom auth headers are supported for enterprise MCP deployments. A ClawHub integration lets users search and install public agent skills directly from the running agent. ## Provider Flexibility nanobot supports a large and growing list of LLM providers including OpenAI, Anthropic, Google Gemini, DeepSeek, Moonshot Kimi, Qwen, MiniMax, Mistral, Xiaomi MiMo, VolcEngine, StepFun, Azure OpenAI, GitHub Copilot, OpenRouter, LM Studio, Hugging Face, vLLM, and Ollama. The provider abstraction was refactored so that adding a new LLM takes only two steps, and the project recently dropped litellm in favor of native openai and anthropic SDKs for stronger reliability. ## Production-Ready Runtime Long-running stability is a primary focus. The runtime hardens user-message persistence, supports mid-turn follow-up injection, auto-compacts long sessions on the fly, and survives restarts thread-aware. A production-ready sandbox, programming Agent SDK, virtual tool-call heartbeat, prompt cache optimization, and macOS LaunchAgent deployment all ship in the box. Windows and Python 3.14 are now first-class supported targets. ## Limitations The MIT-licensed codebase is intentionally minimal, so teams that want heavy enterprise governance features will need to layer their own controls on top. nanobot is positioned as a personal agent rather than a multi-tenant SaaS platform, so per-user isolation and managed control planes are not the focus. The fast release cadence (multiple updates per week) is great for hobbyists but means production deployments should pin specific versions and review release notes carefully before upgrading.