Pick NanoClaw if you are comfortable in a terminal and want an assistant you host yourself, with source small enough to read. It stands in for a cloud chatbot on personal tasks, but it is not a team product: it needs an always-on machine with 4 GB of memory and an AI account you pay for separately.
What changed
- · Release NanoClaw 2026.10.1 second release candidate lists breaking changes (v2026.10.1-rc.2) source
NanoClaw is an open-source personal AI agent that runs on your own Mac or Linux computer and answers through the terminal or messaging apps. Every active session runs in its own Docker container. The repository has passed 31,000 GitHub stars, and Claude is the default provider, with Codex and OpenCode as alternatives you can switch to.
Maker: NanoCo · Autonomy: semi autonomous · Maturity: GA
Underlying models: Claude (default, via the Claude Agent SDK), OpenAI Codex (ChatGPT subscription or OpenAI API key), OpenCode (provider skill)
About NanoClaw
You install NanoClaw on your own Mac or Linux computer, then talk to it from the terminal or from a messaging app such as Telegram, WhatsApp, Slack or Discord; it is open source and built for one person's tasks. Gavriel Cohen created it as a smaller, more contained alternative to OpenClaw, and the company behind it, NanoCo, announced a $12 million seed round led by Valley Capital Partners in May 2026, with Docker, Vercel, Monday.com and Slow Ventures taking part. The code stays under the MIT licence, and NanoCo earns money by selling implementation help to businesses rather than by charging for the software.
A single Node process on your machine receives each message, routes it to an agent group and starts a Docker container for the active session. Inside that container the agent runs on the Claude Agent SDK by default, so Anthropic's models do the thinking, and Codex or OpenCode can replace them for a given group. Each group keeps its own instructions, memory and allowed folders. Credentials sit in a vault called OneCLI, which adds keys to outbound requests so the agent never holds a raw API key by default. That is what makes it an agent and not a chatbot: it can run shell commands, read files and browse the web, inside a boundary you set.
Most people use it for recurring briefings through scheduled tasks, web research, and a memory of their preferences that carries across conversations. The docs also describe prebuilt assistants for jobs such as household planning, competitor research and sales support, and a guide for running a small team of agents. It suits someone comfortable in a terminal who wants an assistant whose source they can read. Anyone who would rather not host a thing can look at the hosted teammate Ako, and Muse is a consumer agent that lives inside Meta's apps. Work that is mostly code belongs with Claude Code, while Hermes Agent is the open-source alternative that tries to learn your workflows over time. The cloud-hosted counterparts come from large vendors: OpenAI runs Dots alongside ChatGPT Work, xAI offers always-on teammates in Grok Bot, Google ships Gemini Spark for Gmail and Calendar, and Manus sells Cue with its own phone and email. The full list sits in the general-purpose agents directory.
The software is free, but the AI behind it is not, and neither is the computer that has to stay switched on. You bring your own credential for Claude or for Codex, which means a subscription or an API key from Claude or OpenAI, and you pay that provider directly. Windows users run it through WSL2, the Linux layer that ships with Windows.
The project moves quickly. Stable release 2026.10.0 is current, and the second release candidate for 2026.10.1 appeared on 22 October 2026 with several breaking changes in its notes, so an update is something to read before you run it. The security documentation is unusually frank that a sandbox is one layer and not a guarantee, which is a useful signal about how the maintainers think.
Pricing
The NanoClaw software is free under the MIT licence. Your AI provider bills you directly, through Claude or through Codex, and you supply the computer it runs on. NanoCo sells implementation services to companies and publishes no rate for them. Self-hosted rivals with the same cost shape include [OpenClaw](/hub/agents/openclaw) and [Hermes Agent](/hub/agents/hermes-agent).
Key Features
- One container per session: Each active session starts in its own Docker container, run as an unprivileged user with all Linux capabilities dropped, so an agent only reaches the folders mounted into it.
- Credential vault: Outbound requests pass through OneCLI's Agent Vault, which adds keys at request time, so by default the agent never holds a raw API key.
- Swappable AI provider: Claude through the Claude Agent SDK is the default, and Codex or OpenCode can replace it for a single agent group without touching the others.
- Channels you choose: Connects to WhatsApp, Telegram, Slack, Discord, Signal and Gmail, and you add only the channels you use as skills, leaving the rest out of your fork.
- Scheduled tasks: Runs jobs on a timetable, such as a morning briefing or a weekly review, and messages the result back to your chat.
- Memory per agent group: A work assistant and a household assistant stay apart, because every group has separate notes, history and folder access.
- Templates and agent teams: Offers prebuilt assistants for jobs like household planning, competitor research and sales support, plus a guide for running several agents together.
Strengths
- The design assumes prompt injection will eventually succeed and then limits what a compromised agent can reach, a franker security stance than most assistants publish.
- An MIT licence and about 12.7k forks mean you can take the code and have a coding agent reshape it to your own needs.
- A funded maintainer with Docker and Vercel partnerships lowers the risk of abandonment, which matters for software you give shell access to.
- Provider and channel choices are not tied to one vendor, so you can change either without leaving the project.
Weaknesses
- Containers limit risk but do not remove it: every session in an agent group mounts the same workspace read-write, so a prompt-injected session can reach that group's files and memory, and the docs say a sandbox is not a guarantee.
- Setup is aimed at technical users: Mac or Linux only, Windows needs WSL2, and the host machine must stay on for the assistant to answer.
- Updates move fast and break things: the 2026.10.1 release candidates list several breaking changes, including a OneCLI gateway upgrade and a Telegram fix, so each update needs its notes read first.
- No benchmarks are published, there is no hosted version for individuals, and the enterprise service has no public price.
Frequently Asked Questions
How much do you pay for NanoClaw?
The software costs nothing: it is MIT-licensed and the whole project sits on GitHub. What you pay is your AI provider, either a Claude subscription or an Anthropic API key billed by usage, or ChatGPT and OpenAI credentials when you pick Codex, plus the electricity and uptime of the computer you run it on. NanoCo also sells implementation help to companies and publishes no rate for it.
Can you use NanoClaw without paying?
Yes for the software, with one catch: it has no AI of its own, so a working install needs a credential for an AI provider, and that is where a bill appears. A Telegram bot costs nothing to create, while WhatsApp and workspace apps such as Slack may need approval or accounts you already hold. Windows users also have to set up WSL2 before the installer will run.
What are NanoClaw's closest competitors?
OpenClaw is the larger self-hosted gateway that NanoClaw grew up as a reaction to, so it fits better when you want a bigger starting set of integrations. Hermes Agent is the better match for an agent that improves from your own workflows over time. Ako is the choice for a team that wants a hosted teammate in its chat apps with no server to look after.
Is NanoClaw better than OpenClaw?
NanoClaw trades breadth for containment. NanoCo's own site lists about 200 runtime TypeScript files against 3,680 for OpenClaw, and describes container isolation against application-level checks, though those figures are the vendor's and not independently verified. It wins on auditability and a hard boundary around the agent, and loses on built-in breadth, so the answer depends on which of those you need.
What does it take to start using NanoClaw?
You need a Mac or Linux computer with at least 4 GB of memory, Git and an AI account; the installer adds Node, Docker and the credential vault for you. Clone the repository, run the setup script, sign in to your AI provider and send your first message in the terminal. The docs suggest setting aside 30 to 60 minutes, and longer if you connect a messaging app.
Top Alternatives
- OpenClaw: NanoClaw gives a smaller codebase and container isolation by default, whereas OpenClaw gives more integrations from the first day.
- Hermes Agent: NanoClaw is the contained assistant you can read end to end, while Hermes Agent is the one that learns your routines as you use it.
- Ako: Self-hosting your own agent points to NanoClaw; a team that prefers a managed teammate in its chat apps points to Ako.
- Claude Code: A personal assistant that answers in chat apps is NanoClaw's job; software work inside a repository belongs to Claude Code.