OpenClaw: The Free AI Assistant You Text From WhatsApp

OpenClaw is one of the most-starred projects on GitHub, and it does not look like a product at all. There is no app store listing to download, no monthly plan to buy, and no company dashboard where your conversations live. You install it on your own computer, link your WhatsApp, and text it like you would text a friend. It reads your email, manages your calendar, edits files, and runs commands on the machine it lives on.

As of October 27, 2026, the repository at github.com/openclaw/openclaw has 393,214 stars and 82,665 forks. It passed React and the Linux kernel in star count back in March, according to third-party tracking of the project. That is a strange position for a personal assistant with no marketing department, so it is worth asking what OpenClaw actually is, what it costs in practice, and where the sharp edges are. The answer to the last question matters more than the first two, and OpenClaw's own documentation says so in plain language.

What OpenClaw actually is

OpenClaw is an open-source personal AI assistant that runs on your own hardware. The core piece is called the Gateway: a local control plane that holds sessions, tools, events, and channel connections. You reach it through the messaging apps you already use. The README lists Discord, iMessage, Slack, Teams, Telegram, and WhatsApp, plus more than 20 channels in total, and native apps for macOS, iOS, Android, Windows, and Linux. One Gateway can run as a personal assistant on a laptop or as a shared team deployment; configuration is the only difference.

The model is a plugin, not the product. OpenClaw works with Claude, Codex, local models through Ollama, and other hosted providers. Swap the model and the rest of the setup stays the same. That is the design decision that separates it from the hosted assistants it competes with: your state, memory, and credentials live on your hardware, and the project says it phones home for nothing but a daily version check. Anonymous feature statistics are opt-in. Once connected, it triages an inbox, manages a calendar, researches topics on the web, reads and writes files, and runs shell commands. Skills and plugins extend it through a community registry called ClawHub, and companion apps add voice, camera, and device-local actions on supported platforms.

The star count, and the story behind it

OpenClaw did not start under this name. It appeared in late November 2025 as Warelay, a WhatsApp gateway, then became Clawdbot, then Moltbot after Anthropic asked for a name change over its trademark in January 2026. It has been OpenClaw since January 30, 2026. The creator is Peter Steinberger, an Austrian developer who built it for Molty, his space-lobster AI assistant. Steinberger works at OpenAI, which confuses people: OpenAI is also one of the project's donors, alongside Amazon, Microsoft, Nvidia, and Red Hat, but the project is stewarded by the OpenClaw Foundation, an independent 501(c)(3) nonprofit set up in July 2026. The README is explicit that donors do not own or direct the project.

The growth has been extreme by any measure. Third-party tracking put the count at about 392,000 stars on October 20, 2026, and the live GitHub count on October 27 was 393,214. The foundation said in July that 4.5 million new installs were being set up every week, a figure reported by independent coverage rather than a number I can verify from the repository itself.

Stars are not a quality metric, and a star count this large can mislead. What it does measure is distribution: a lot of people have heard of OpenClaw, and a lot of them installed it. Whether they kept running it is a different question, and the project does not publish retention numbers. Treat the 393,000 as a signal of attention, not of reliability.

How it works from WhatsApp

WhatsApp is the headline channel, and the setup is specific. OpenClaw connects through WhatsApp Web using the Baileys library, which means the Gateway owns a linked WhatsApp session rather than going through Meta's official Business API. You add the channel during onboarding, or later with openclaw channels add --channel whatsapp, and log in with openclaw channels login --channel whatsapp. The plugin installs on demand from the npm package @openclaw/whatsapp.

Access control is where the design shows its priorities. Direct messages use a pairing policy by default: unknown senders are not automatically paired, and approving one takes an explicit command, openclaw pairing approve, with the channel and a code. You can switch to an allowlist of phone numbers, or to an open policy, but the documentation does not recommend open for personal deployments. Group chats get two layers: a group membership allowlist, and a sender policy that decides who inside an allowed group can trigger the assistant. Group replies require a mention of the bot by default. An assistant that can read your files and send your email should not answer every stranger who messages it, and OpenClaw's defaults are built around that assumption.

How to set it up

The installer supports macOS, Linux, and Windows, and provisions a supported Node.js runtime when needed. On macOS, Linux, or WSL2: curl -fsSL https://openclaw.ai/install.sh | bash. On Windows PowerShell: iwr -useb https://openclaw.ai/install.ps1 | iex. If you already manage Node.js, install the package with npm install -g openclaw@latest --allow-scripts=openclaw, which needs Node 24.22.1 or later, or Node 26.12.1 or later, with Node 26 recommended.

After installing, run openclaw onboard --install-daemon. Onboarding verifies model access, creates the workspace, and configures the Gateway. You will need an API key from a model provider, or a Claude or ChatGPT subscription to sign in with. Then openclaw gateway status checks the Gateway, and openclaw dashboard opens the Control UI, where you send a first message to confirm the assistant is working.

Two practical notes before you start. First, OpenClaw needs a machine that stays on. A laptop that sleeps every night will not answer your 2 a.m. message. A Mac mini is the popular always-on host, but any Mac, Linux machine, Windows PC, small cloud server, or Raspberry Pi-class box works; you only need a Mac for macOS-only features such as iMessage. Second, WhatsApp linking uses your own number through WhatsApp Web. That is convenient, and it is also why the security section below exists. Independent setup guides put a first install at 15 to 30 minutes; the docs describe the Gateway itself as about five minutes, before channel setup.

What it actually costs

OpenClaw itself costs nothing. The code is MIT-licensed, there is no paid tier, no hosted service, and no token to buy. The bill arrives one step downstream, from the model provider you point it at. There are three realistic ways to pay for that, and they land very differently.

The cheapest route for most people is a subscription you already pay for. OpenClaw supports signing in with a ChatGPT or Codex subscription, and OpenAI explicitly allows that use. It also works with a Claude subscription. Claude Pro costs $20 a month billed monthly, or $17 a month when billed annually at $200 up front, and includes Claude Code with the same usage pool. Claude Max starts at $100 a month and offers five or twenty times Pro's usage, plus $100 to $200 a month in API credits. Prices exclude tax, and usage limits apply on a rolling five-hour window with weekly caps on top.

The second route is pay-as-you-go API usage. On Anthropic's published rate card, Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens, with cache reads at 5 percent of the input price. Claude Haiku 5.5, the budget option for high-volume work, costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, rising to $0.50 and $2.50 above that line. A light personal assistant that mostly answers short messages will cost a few dollars a month this way. A heavy one that reads long email threads and browses the web all day will cost considerably more, and the bill scales with output tokens, which are five times the input price on Sonnet.

The third route is local models through Ollama, where the model itself is free and the cost is electricity and hardware you already own. This is the genuinely zero-marginal-cost option, and also the one with the weakest models. A small local model will handle simple triage and short answers. It will not match Claude or GPT on long, multi-step tasks.

There is no free lunch in the subscription route either. A $20 Claude Pro subscription shared between the Claude app, Claude Code, and an always-on OpenClaw gateway draws from one usage pool, and an assistant that runs all day will hit the weekly caps faster than a person typing at a keyboard.

The security part you cannot skip

This is the section that decides whether OpenClaw is a good idea for you, and the project's own README opens it with a warning worth quoting in full: "Treat inbound messages as untrusted input. DM-capable channels pair unknown senders by default; approve a pairing request with openclaw pairing approve. Tools run on the host for the main session unless you configure sandboxing. Read the security guide, exposure runbook, and sandboxing guide before connecting other users or exposing the Gateway remotely."

By default, the tools in your main session run directly on your host machine, with no sandbox between the model and your files. That is what makes OpenClaw powerful: it can actually do things, not just talk about them. It is also what makes it dangerous. A prompt-injection message in a group chat, a malicious link summary, or a pasted instruction from a stranger can, in the worst case, become a command that runs on your computer. The pairing policy and the allowlist exist to narrow that window. Sandboxing exists to close it further. Skipping both because the install worked is the single most common mistake a new OpenClaw user can make.

The supply chain deserves the same care. Skills and plugins from ClawHub are third-party code, and installing one hands it the same tools your assistant has. ClawHub now runs a security review program for uploaded skills and plugins, according to independent coverage of the ecosystem, but a review program is not a guarantee. Install skills from authors you trust, and read what a skill can do before you enable it.

None of this makes OpenClaw unsafe to run. It makes it unsafe to run carelessly. The difference is the entire product: a hosted assistant absorbs these risks for you and charges a subscription for the privilege. OpenClaw hands you the risk, the controls, and the bill for the model, and charges nothing for the software. If nobody on your side will read the security guide before connecting a real inbox, do not install it yet.

OpenClaw vs the alternatives

The honest comparison is not against other open-source projects. It is against the hosted assistants most people already have, and against the other free assistant that made news this month.

OpenClaw wins on control and on breadth of channels, and it is the only one of the four that meets you inside WhatsApp. It loses on convenience: there is no hosted fallback, no support desk, and no one to call when the Gateway stops answering at midnight. Wabi, by contrast, is free and polished and runs in the cloud, but your data sits on someone else's servers and there is no public API. The ChatGPT and Claude apps are the safe defaults for most people, and they stay inside their own apps for a reason: the company, not you, carries the security risk.

For a fuller picture of the hosted text-agent field, our roundup of 7 AI agents you can text in 2026 covers seven services with their prices and privacy terms. Wabi, the free family assistant in that field, gets a closer look in our Wabi 2.0 review. And if the coding side of OpenClaw's model plugins interests you, the 7 best AI coding agents in 2026 roundup covers the agents those plugins wrap.

Pros and cons

Pros: genuinely free software under a permissive MIT license, with no paid tier and no hosted service to lock you in; runs on your own hardware, so your state, memory, and credentials stay with you; meets you in WhatsApp, Telegram, Slack, Discord, Signal, and iMessage instead of one more app; models are swappable plugins, so you are not tied to one provider; a large, active community and a skills registry in ClawHub; an independent nonprofit foundation, not a venture-backed company, governs it.

Cons: tools run on your host by default unless you configure sandboxing, which is a real security risk for careless setups; needs an always-on machine, so a sleeping laptop means a silent assistant; setup takes 15 to 30 minutes and real configuration after that; model costs are yours, and a busy assistant can burn through a subscription's usage caps; no support desk and no hosted fallback when something breaks; third-party skills are a supply chain you inherit.

FAQ

Is OpenClaw really free?

The software is free under the MIT license, with no paid tier, no hosted service, and no token to buy. You pay for the AI model behind it: a Claude or ChatGPT subscription, pay-as-you-go API usage, or a free local model through Ollama. The software being free and the assistant being free are two different statements, and only the first one is true.

Is OpenClaw owned by OpenAI?

No. It is stewarded by the OpenClaw Foundation, an independent 501(c)(3) nonprofit. Its creator, Peter Steinberger, works at OpenAI, and OpenAI is one of more than 30 donors, alongside Amazon, Microsoft, and Nvidia. The README states that donors do not own or direct the project.

Can I use OpenClaw with WhatsApp?

Yes. WhatsApp is one of the headline channels, connected through WhatsApp Web using the Baileys library. You link it during onboarding or with openclaw channels login --channel whatsapp. Direct messages use a pairing policy by default, so unknown senders are not automatically paired.

Is OpenClaw safe to install?

It is safe to install carefully and unsafe to install carelessly. Treat inbound messages as untrusted input, keep the pairing or allowlist policy in place, read the security guide before connecting other people, and configure sandboxing so tools do not run directly on your host. Third-party skills from ClawHub are code you inherit, so install them from authors you trust.

Do I need a powerful computer to run OpenClaw?

No. The Gateway itself is light; the heavy work happens at the model provider. Any Mac, Linux machine, Windows PC, small cloud server, or Raspberry Pi-class box works as an always-on host. You only need a Mac for macOS-only features such as iMessage. The real requirement is uptime, not horsepower.

Key takeaways

• OpenClaw is a free, open-source AI assistant that runs on your own computer and answers you inside WhatsApp, Telegram, and 20-plus other chat apps.
• It holds 393,214 GitHub stars as of October 27, 2026, making it one of the most-starred open-source projects ever, but stars measure attention, not reliability.
• The software costs nothing under its MIT license; you pay for the model, through a Claude or ChatGPT subscription, API usage, or a free local model.
• WhatsApp works through WhatsApp Web, with a pairing policy that keeps unknown senders out by default, and setup takes 15 to 30 minutes.
• Its own README warns you to treat inbound messages as untrusted input and set up sandboxing before connecting other people, and that warning is the whole story.

OpenClaw is the rare open-source project that lived up to its own hype in one specific way: it actually does the thing it promises, which is putting a capable AI assistant inside the chat apps you already live in, on hardware you own. The price of that is not money. It is the half hour of setup, the always-on machine, and the security reading that its own documentation insists on. Do that reading first, and the lobster earns its keep. Skip it, and you have handed the keys to your computer to the internet.

Sources: OpenClaw GitHub repository and README (github.com/openclaw/openclaw); OpenClaw documentation, getting started and WhatsApp channel guides (docs.openclaw.ai); Claude pricing and API pricing pages (claude.com/pricing, platform.claude.com/docs/en/about-claude/pricing), read October 28, 2026; independent coverage from madrobot.blog, ayautomate.com, rerun.build, frontiernow.dev, and Medium. Star count verified live on the GitHub API on October 27, 2026.