nanobot

Knock, knock. 🐈nanobot v0.2.0🥳is out. Your open-source personal AI agent just shipped a major upgrade.

Three months ago, nanobot showed up on Hacker News as a 4k-line Python program with three chat channels and every model request routed through a LiteLLM proxy. Someone called it "the irreducible core of an agent." Someone else called it a toy. Three months later, it's still here. Let's talk about what got cut first. LiteLLM is gone. Entirely. In March, LiteLLM 1.82.7 and 1.82.8 on PyPI were compromised, a supply chain attack that made a lot of people nervous. We ripped it out and replaced the whole thing with three SDKs: openai, anthropic, and httpx. This wasn't some grand architectural vision. We were scared. But the side effect was surprisingly good: fewer dependencies, streaming that actually works, reasoning tokens that stop leaking into chat. Now what grew. 14 chat channels (from 3, added Discord, Slack, WeChat, Teams, DingTalk, QQ, Email, Matrix, WebSocket, WeCom, MoChat). 6 native LLM providers (no proxy anymore, each talking directly to the upstream API: OpenAI, Anthropic, Azure, Bedrock, GitHub Copilot, Codex, plus an OpenAI-compat layer for everything else). MCP support. A built-in WebUI that ships inside the Python wheel (no separate deploy). Bubblewrap sandboxing in Docker. Sustained goals that survive context compaction. What didn't change. The core loop is still: receive message, build context, call LLM, execute tools, respond. One agent, one loop. No orchestrator, no planner, no multi-agent framework. runner.py is the entry point, 1,308 lines, readable in about 30 minutes if you want to understand how a request flows from channel to provider and back. The numbers: 4k lines to 45k lines of Python(though the core runtime is under 8k). 1 test file to 204. 31 contributors to 340. Version 0.1.3 to 0.2.0. 42K stars. Here's the honest part. 42K stars is a lot. Most of that traffic came from the February HN post. The tagline compares it to OpenClaw, which is both the project's best marketing and its biggest misrepresentation. OpenClaw works. So does nanobot. The difference is that nanobot's code is simple enough to read in an afternoon and modify the next day. 45k lines total, but the irreducible core is under 8k. The rest is scaffolding you can ignore. One loop, no magic. It's less a product and more an SDK. In an era where LLMs can write code for you, what matters isn't features, it's having a solid foundation you actually understand and can shape to fit your own needs. Code is cheap. Understanding is not. MIT licensed, Python 3.11+.

Built a Minecraft-style FPS game with an open-source AI agent. Cost me $0.50. nanobot's new /goal command holds a sustained objective across turns — I just set the goal once, walked away, came back to a working game. DeepSeek-V4-Pro from @deepseek_ai doing the heavy lifting. Short demo ↓

/goal build super mario game in @nanobot_project! End-to-end recording with glm-5.1 from @Zai_org. It only cost $0.1 in one-shot prompt with our lightweight agent! Have fun!

✦ /goal — sustained objectives across turns, no more one-shot replies ✦ image generation, end to end ✦ WebUI now ships inside the wheel ✦ 5 new providers + fallback_models ✦ a real agent-loop refactor 105 PRs · 33 contributors · 20 newcomers → github.com/HKUDS/nanobot/…

Knock, knock. 🐈nanobot v0.2.0🥳is out. Your open-source personal AI agent just shipped a major upgrade.

Happy Friday ☀️ Anyone waiting for the next release? 👀

nanobot has been officially featured by DeepSeek as one of the "Awesome DeepSeek Agents"! 🚀 We're thrilled to be featured alongside popular tools like Claude Code, OpenCode, OpenClaw, and GitHub Copilot, in their curated list of 15 recommended AI agents that integrate seamlessly with DeepSeek models. Grateful for this recognition and excited to continue improving nanobot to serve users across different fields. Thank you to everyone who has supported us on this journey! Try nanobot: github.com/HKUDS/nanobot Check out the full list: github.com/deepseek-ai/aw… #nanobot #AIAgent

Thank you @deepseek_ai 🎉

nanobot v0.1.5.post3 is here🎉

🐈#nanobot v0.1.5.post3 is here 🎉

🚀 Big milestone! nanobot just ranked TOP-4 globally for Personal Agents on @OpenRouter! Big congrats to @xubinrencs for his hard work and dedication!!

#nanobot is on the rise 🚀 We’re ranked Top-4 in both Productivity and Personal Agents by @OpenRouter.

🐈 nanobot v0.1.5.post2 release 🚀 👉 Windows support - nanobot now runs much more smoothly on Windows, with Python 3.14 support and CI coverage to back it up. Native Windows usage is no longer an afterthought. 👉 read_file now understands Office docs - DOCX, XLSX, and PPTX are now readable directly by the agent. Real-world files go in without manual conversion. 👉 OpenAI-compatible SSE streaming - the API now streams with proper SSE for /v1/chat/completions. If your app already works with OpenAI-style clients, plugging into nanobot gets much easier. 👉 More integrations - Microsoft Teams support lands, LM Studio works with nullable API keys, MiniMax gets thinking-mode support, and MyTool gives safer runtime self-inspection. 👉 Reliability upgrades across the stack - session files are safer, memory recovery is tougher, cron is quieter, provider fallbacks are sturdier, and Telegram/email handling is more polished. Less weirdness, more trust. 👉 WebUI is taking shape - a dedicated browser UI with WebSocket chat, i18n, typography polish, and dark mode is now in the repo. Still source-only for now, but it's coming together fast. Try it out: pip install --upgrade nanobot-ai Learn more in the v0.1.5.post2 release notes: github.com/HKUDS/nanobot/…

🐈 nanobot v0.1.5.post1 release 🚀 👉 Mid-turn injection - send a follow-up while the agent is still working. It folds into the active turn instead of waiting behind a lock. On streaming channels the reply keeps flowing in the same response. Long tool-call chains finally feel like a conversation. 👉 Dream skill discovery - Dream now spots repeated workflows and promotes them into reusable skills under skills/. Your agent's capabilities grow from how you actually use it — no manual skill authoring needed. 👉 WebSocket channel - a new first-class channel that runs a WebSocket server with streaming, token auth, and TLS. Web apps, CLIs, Chrome extensions, scripts — anything that opens a socket can talk to the agent in real time. 👉 Deeper channels everywhere - Telegram location pins, Discord streaming + proxy, Feishu done-emoji + Lark global, QQ & WeCom full media, Slack named targets, API file uploads with automatic PDF/DOCX/XLSX/PPTX extraction across all channels. 👉 Smarter reasoning - Anthropic adaptive thinking lets Claude decide when and how much to think. Kimi thinking models get automatic reasoning injection via config. MCP resources and prompts are now exposed as tools the agent can query directly. Try it out: pip install --upgrade nanobot-ai Learn more in the v0.1.5.post1 release notes: github.com/HKUDS/nanobot/…

🐈nanobot fully supports the Agent Skills format and is listed at Agent Skills Client Showcase (agentskills.io/clients). This is yet another step towards a standardized and perfect agent ;)

🐈 nanobot v0.1.5 — the release where it stops being a prototype. 🏗️ Long-running reliability — structured retry classification, context consolidation, checkpoint recovery 🧠 Dream memory — two-stage consolidation that distills conversations into lasting knowledge, git-tracked ⚡ Programmable kernel — Python SDK, OpenAI-compatible API, composable lifecycle hooks 🔐 Production hardening — bwrap sandboxing, non-root containers, env var interpolation for secrets 🌐 7 new providers — GPT-5, MiMo, Qianfan, Langfuse integration, and more 57 PRs, 27 new contributors. Docs: nanobot.wiki pip install -U nanobot-ai