OpenClaw creator Peter Steinberger (right) with co-host Tomas Taylor backstage at the first ClawCon, at Frontier Tower in San Francisco, on February 4, 2026. Image: LogicFlow99 / Wikimedia Commons, CC0 1.0, cropped
OpenClaw is a free, open-source AI assistant that you run on your own computer and talk to through the chat apps you already use, such as WhatsApp, Telegram, Discord, Slack and iMessage. You plug in an AI model of your choice (Claude, ChatGPT or a local one) and it can read and write files, browse the web, run commands, keep a memory and work through tasks on a schedule. Austrian developer Peter Steinberger built it, mostly by directing AI coding agents, and with about 392,000 stars it is now one of the most-starred projects on GitHub. Here is what it is, what it costs, how to install it, how to connect Claude, ChatGPT or Ollama, and whether it is safe.
Jump to:
[What is OpenClaw?](#what)
[Is it free?](#free)
[How to install](#install)
[Claude, ChatGPT and Ollama](#models)
[Is it safe?](#safe)
[vs Claude Code and Hermes](#vs)
[Where it works](#where)
[FAQ](#faq)
What is OpenClaw and what is it used for? #
OpenClaw’s own FAQ calls it “an AI assistant you run on your own infrastructure”. The core is the Gateway, a program that runs all the time on your laptop, a Mac mini or a cloud server. It connects your chosen AI model to your chat apps, your files and tools such as a web browser, and keeps your settings, memory and logins on your own hardware. You message it like a friend, and it replies and acts.
The project lists its top everyday uses as:
- Personal briefings: summaries of your inbox, calendar and news.
- Research and drafting: quick research and first drafts of emails or documents.
- Reminders and follow-ups on a schedule.
- Browser automation: filling in forms and repeating web tasks.
- Cross-device jobs: send a task from your phone, let a server run it, get the answer back in chat.
Extra abilities come as skills and plugins, many shared on its marketplace, ClawHub. There are companion apps for macOS, iOS, Android and Windows, and teams can share one Gateway.
The name has a history. According to the project’s own account, it began as Warelay, a WhatsApp gateway, then became Clawdbot. Anthropic asked for a new name over its trademark in January 2026, so it became Moltbot for a few days and then OpenClaw on January 30, 2026. Since July 2026 it has been run by the OpenClaw Foundation, a US non-profit. Steinberger joined OpenAI earlier in 2026, and OpenAI is a donor alongside Amazon, Microsoft, Nvidia, GitHub, Red Hat, Tencent and others, but the project says no donor owns or directs it.
Is OpenClaw free? What it costs to run #
The software is free. OpenClaw is MIT-licensed, and the foundation says it has “no paid tier, hosted service, or token”. Your costs come from two places:
| Cost | What you pay |
|---|---|
| OpenClaw itself | Free |
| Computer to run it | Any machine you already own, or a small cloud server. Minimum 1 vCPU, 1GB RAM and about 500MB of disk; 2GB or more recommended |
| Anthropic or OpenAI API key | Pay as you go, per token, at the provider’s API prices |
| Claude Pro or Max, ChatGPT Plus or Pro | Counts against your plan’s usage limits |
| Local model (Ollama, LM Studio, llama.cpp) | Free, but needs a capable computer |
The model bill is the one to watch: an assistant that runs all day, checks your inbox and browses the web uses far more tokens than a chat window. The command openclaw models status shows the usage windows your provider reports.
How to install OpenClaw on Mac, Windows and Linux #
OpenClaw’s install guide offers three routes:
- Desktop apps: on Windows, download the signedWindows Hub installer (x64 or Arm); on a Mac, download the
.dmgfromOpenClaw’s GitHub releases and open OpenClaw.app. Both can set up a local Gateway on first run. - Installer script (the fastest route). On macOS, Linux or WSL2:
curl -fsSL https://openclaw.ai/install.sh | bash. In Windows PowerShell:iwr -useb https://openclaw.ai/install.ps1 | iex. It installs Node.js if needed and starts the setup wizard. - npm , if you already manage Node.js (version 24.16 or later, or 26.1 or later):
npm install -g openclaw@latest, thenopenclaw onboard --install-daemon.
Onboarding checks your model works, creates a workspace and sets up the Gateway. Then run openclaw dashboard to open the control panel and send a test message. Docker, Nix and cloud deployments are covered in the docs.
Not working? Start with openclaw status, then openclaw doctor, which repairs common config problems, and openclaw logs --follow. On Windows, “git not found” means you need Git for Windows, and “openclaw is not recognized” means npm’s folder (usually %AppData%\npm) isn’t on your PATH; reopen PowerShell after fixing either.
You don’t need to buy a Mac mini. It is a popular always-on host, but the docs say a small cloud server, home server or Raspberry Pi-class box works too. You only need a Mac for macOS-only features such as iMessage.
How to use OpenClaw with Claude, ChatGPT or Ollama #
OpenClaw works with most model providers, and you choose during onboarding or later with openclaw onboard.
- Claude subscription: OpenClaw can reuse aClaude Code login on the same computer for Pro, Max, Team and Enterprise plans. ItsAnthropic page warns that this usage comes out of your plan’s limits rather than a separate allowance, and that Anthropic d a promised monthly Agent SDK credit on June 15, 2026. For an always-on or shared assistant, OpenClaw recommends an Anthropic API key instead.
- ChatGPT or Codex subscription: fully supported, and OpenClaw says OpenAI explicitly allows it. OpenClaw was also one of the partner tools in the“Sign in with ChatGPT” scheme OpenAI announced at DevDay , which lets Plus and Pro users spend their plan allowance there.
- Ollama: run
openclaw onboard, pick Ollama and choose Cloud + Local, Local only or Ollama Cloud. Use the native addresshttp://localhost:11434, not the/v1one, whichbreaks tool calls . Open models such asMeta’s Muse Glimmer are built for this kind of local agent. - Others: Google (AI Studio or Vertex API key), OpenRouter, AWS Bedrock, MiniMax, Qwen and Z.AI coding plans, and LM Studio.
OpenClaw’s own advice on small local models is blunt: “Usually no.” It says they cut off context and lack hosted safety filters, which raises the risk of prompt injection, so run the largest local model your machine can handle.
Is OpenClaw safe? #
OpenClaw is reasonably safe with its default settings, and much riskier once you open it up. Out of the box the Gateway only listens on your own machine, and most chat channels answer a stranger with a pairing code instead of passing their message to the AI. You approve each new contact with openclaw pairing approve.
The risks the project’s security guide flags are:
- Exposed Gateways: making it reachable from the internet without a password or token. Use Tailscale, an SSH tunnel or token auth instead.
- Prompt injection: any web page, email or document it reads can carry hidden instructions. Keep web browsing away from agents that can also run commands, and use a read-only agent to summarise untrusted content.
- Third-party skills: treat them as untrusted code. ClawHub shows a scan result for each skill before you install, but OpenClaw says scans “are not a complete security boundary”. Read a skill before enabling it and prefer known authors.
- No sandbox by default: tools run directly on your computer unless you turn on sandboxing.
Run openclaw security audit --deep before you connect anyone else. Microsoft’s new Windows agent sandbox already supports OpenClaw, and Nvidia’s NemoClaw adds its OpenShell sandboxing. Remember who is responsible: anything your agent sends or buys is done in your name, an issue Grok’s owner spelled out when it made users liable for its agents.
OpenClaw vs Claude Code, Codex and Hermes Agent #
| OpenClaw | Claude Code / Codex | Hermes Agent | |
|---|---|---|---|
| What it is | Always-on personal assistant | Coding agent for a project folder | Always-on personal assistant |
| Price | Free (pay for the model) | Subscription or API | Free software, optional paid tiers |
| Chat apps | 20+ | No (terminal, IDE, app) | Yes |
| Who runs it | Non-profit foundation | Anthropic / OpenAI | Nous Research (venture-funded) |
| GitHub stars (Oct 11, 2026) | About 392,000 | n/a | About 253,000 |
Scroll sideways to see the full table.
OpenClaw describes itself as “an assistant and coordination layer, not an IDE replacement”: for the fastest coding inside a project, use Claude Code or Codex, and use OpenClaw for memory, reminders and reaching your tools from your phone. In fact OpenClaw can run Codex inside it as a plugin. Its closest rival is Hermes Agent; OpenClaw’s own comparison stresses that it is donor-funded while Hermes’s maker sells paid tiers, but it also admits that its sandboxing is off by default.
Where OpenClaw works #
OpenClaw has no country restrictions: it is software you download and run, so it works anywhere you can install it, including the US, UK, Canada, Australia, Germany, the rest of the EU and India. The limits come from the model you plug in. Claude and ChatGPT are only available in their supported countries, and neither is officially offered in mainland China, where people tend to use Qwen, MiniMax or Z.AI plans or a local model instead. OpenClaw also lets you keep hosted model traffic in a chosen region. With a local model, nothing has to leave your machine at all.
Last checked: October 11, 2026, against OpenClaw’s docs and README, its npm releases and GitHub star counts.
Frequently asked questions #
Is OpenClaw free?
Yes. OpenClaw is MIT-licensed open-source software with no paid tier, hosted service or token. What you pay for is the AI model behind it: an API bill, a subscription you already have, or nothing if you run a local model through Ollama or LM Studio.
How much does OpenClaw cost to run?
The software is free and needs only 1 vCPU, 1GB of RAM and about 500MB of disk. Running costs depend on the model: Anthropic and OpenAI API keys bill per token, a Claude or ChatGPT subscription draws from your plan’s limits, and a local model costs only electricity.
Is OpenClaw safe to use?
It is safe enough with its defaults, which keep the Gateway on your own machine and make strangers pair before your bot will answer them. The real risks come from opening it up: exposing the Gateway to the internet, giving it shell or browser access while it reads untrusted web pages or emails, and installing third-party skills without reading them. Run openclaw security audit before you connect anyone else.
Can I use OpenClaw with my Claude subscription?
Yes. OpenClaw can reuse a Claude Code login for Pro, Max, Team and Enterprise plans, but the usage comes out of your plan’s limits; it is not a separate free allowance. OpenClaw recommends an Anthropic API key for always-on or shared setups because the billing is more predictable.
Can I use OpenClaw with ChatGPT?
Yes. OpenClaw fully supports signing in with a ChatGPT or Codex subscription, and OpenAI explicitly allows that use. You can also use an OpenAI API key.
Does OpenClaw work with Ollama?
Yes. Run openclaw onboard, pick Ollama and choose Cloud + Local, Local only or Ollama Cloud. Use the native address (http://localhost:11434) rather than the /v1 one, which breaks tool calls.
Do I need a Mac mini for OpenClaw?
No. A Mac mini is a popular always-on host, but OpenClaw runs on any Mac, Linux machine, Windows PC, small cloud server or Raspberry Pi-class box. You only need a Mac for macOS-only features such as iMessage.
Is OpenClaw owned by OpenAI?
No. OpenClaw is run by the OpenClaw Foundation, an independent US non-profit set up in July 2026. Its creator, Peter Steinberger, works at OpenAI, and OpenAI is one of more than 30 donors, alongside Amazon, Microsoft, Nvidia and others, but donors do not own or direct the project.
Is OpenClaw dead?
No. As of October 11, 2026, it had about 392,000 GitHub stars, version 2026.10.1 came out on October 10 and a new beta followed on October 11. In July the foundation said 4.5 million new installs were being set up every week.
What was OpenClaw called before?
It started as Warelay, a WhatsApp gateway, then became Clawdbot. After Anthropic asked for a name change over its trademark in January 2026, it was briefly Moltbot, then OpenClaw from January 30, 2026.
Sources: OpenClaw on GitHub; OpenClaw FAQ; install guide; Anthropic provider docs; Ollama provider docs; security guide; Why OpenClaw; OpenClaw lore; Introducing the OpenClaw Foundation.