NanoClaw is a small open-source framework for running a personal Claude agent that plugs into the messaging apps you already use. It is maintained by NanoCo AI, written in TypeScript under an MIT license on top of Anthropic's Claude Agent SDK, and it is aimed at developers and self-hosters who want an assistant sitting in their WhatsApp or Slack without granting a large, unfamiliar codebase the run of their machine.
What it does
NanoClaw connects a Claude-powered assistant to WhatsApp, Telegram, Slack, Discord, Gmail and similar apps. Once connected, the assistant can browse the web, manage files and run scheduled tasks for you, so it behaves less like a chat window you visit and more like a background worker you message.
The distinguishing claim is size. The project says its core is roughly 500 lines of TypeScript, "a codebase you can understand in 8 minutes". That is the maintainers' own framing rather than an independent audit, but the code is public and short enough that any reader can check the claim in an afternoon.
How it works
Each conversation, and each agent, runs inside its own isolated Docker container (Apple containers are also supported). The assistant does its work inside that box, so it does not hold direct access to your whole machine or to your raw API keys. That is an operating-system-level boundary rather than an application-level permission list, and it is the exact design difference the project markets against OpenClaw, which relies on allowlists and pairing codes instead.
A few numbers travel with the project, so it is worth saying who produced each one:
- Roughly 500 lines of core TypeScript, against OpenClaw's reported 400,000-plus lines and 70-plus dependencies: the project's own comparison, repeated by VentureBeat.
- About 31,000 GitHub stars and 12,800 forks as of October 15, 2026, on a repository created in late January 2026: figures the README attributes to the GitHub API.
- Creator Gavriel Cohen, a former Wix.com engineer, says he wrote the first version in about 48 hours after discovering OpenClaw had captured every WhatsApp group on his account, and that viral interest turned into a Docker partnership within six weeks. That is his account, given to VentureBeat.
- Docker backed NanoCo's 12 million dollar seed round, and Singapore's foreign minister has publicly called NanoClaw his second brain.
Getting started
You need two things the materials name plainly: a Docker runtime on the machine that will host the agents, and your own Claude API key. The code is free and MIT-licensed; the intelligence is not. Every message your assistant handles is a paid call to Anthropic on your key, and nothing in the project's materials describes a local-model path that would let you skip that bill. Budget accordingly before you wire the agent into a busy group chat.
Beyond Docker and the key, the project does not list hardware requirements: no operating system, chip, memory or disk figures appear in the materials, so treat anything exotic as untested. Installation is a plain clone of the repository at github.com/nanocoai/nanoclaw; there is no separate installer, release binary or app-store build to hunt down, and the materials do not quote a single setup command, so follow the repository's own instructions rather than a one-liner from a video.
When to use it / when not
The main catch is the Docker dependency. NanoClaw's security story is container isolation, which means a container runtime is not optional, and that rules out hosts where you cannot install one. The second catch is age: this repository was created in late January 2026, so it is a young project with a fast-moving commercial parent.
Use it instead of OpenClaw when you want the same job done by a codebase you can read, and when you would rather trust an OS-level sandbox than an application-level allowlist. Don't use it if you just want to ask Claude questions: for that, the official Claude apps cost you nothing extra to set up and need no container at all.
Alternatives
OpenClaw is the obvious comparison and the one most readers already know: the self-hosted AI-agent project, formerly Clawdbot and briefly Moltbot, that passed 200,000 GitHub stars after launching in November 2025. It is far larger and far more featureful, and its permission model is application-level. If you want maximum surface area and an enormous community, that is the one. If what unsettles you is how much a big agent can quietly read, NanoClaw is the counter-proposal.
Take NanoClaw seriously if you are a developer who wants an always-on assistant reading real messages and you are not willing to accept a six-figure line count as the price of entry. The honest trade is clear: you supply Docker and a paid Claude key, and in exchange you get a sandboxed agent whose core you can audit yourself in one sitting. For anyone who got nervous the first time an agent asked for full access to their chat history, that trade is a reasonable one.