--- title: "Install OpenClaw on your Mac" description: "The complete install guide for macOS: the official installer step by step, Ollama's one-command launch, the one-click path with HolaClaw, and the day-2 work of keeping OpenClaw running and updated." updated: 2026-08-19 canonical: https://holaclaw.ai/docs/tutorials/install-openclaw --- There are three good ways to get OpenClaw running on a Mac: the **official installer**, **Ollama's `ollama launch openclaw`** if you want a local model wired up in the same step, and **[HolaClaw](/download)**, our Mac application. The first two options require you to run commands in the terminal and will install OpenClaw directly on your Mac. HolaClaw is a regular Mac app: download it, click, and OpenClaw runs inside an isolated virtual machine, no terminal involved. This guide covers all three, with the manual path in full: requirements, the install step by step, where everything lives, keeping the Gateway running, and updating without breaking your setup. ## Ways to run OpenClaw on a Mac - **The official installer.** OpenClaw's CLI and Gateway on your Mac, running as your user. The default path: [the step-by-step below](#option-1-the-official-installer). - **A package manager.** npm, pnpm, or bun, if you already manage global Node packages. Covered in step 1. - **Ollama.** One command installs OpenClaw and wires a local model as its brain. See [install with one command](#option-2-ollama). - **Docker.** Official images exist (`ghcr.io/openclaw/openclaw`). The docs position them for isolated, throwaway gateways rather than the everyday setup. - **A dedicated Mac.** The same install on a spare, always-on machine, like a mac mini. See [on a dedicated Mac](#on-a-dedicated-mac-like-a-mac-mini). - **HolaClaw.** Our Mac app: OpenClaw inside an isolated VM, one click, no terminal. See [install with HolaClaw](#option-3-install-with-holaclaw-native-mac-app). ## What you need - **A Mac with an M-series chip.** We recommend it for performance, and HolaClaw and local models require it. - **A decision on your AI provider.** An API key or account with a cloud provider (Anthropic, OpenAI, Google, and others), or no key at all with a local model served by Ollama or HolaClaw. [Choosing an AI provider](/docs/ai-providers/choosing-a-provider) walks through the trade-offs. - **Memory, if you go local.** Powering OpenClaw with a local model is memory-bound. [We tested six models](/blog/running-openclaw-with-local-models) to see which Macs handle which. With a cloud provider, any Mac that runs Node is enough. ## Where we tested this guide We ran this guide on the two MacBook Airs we test with: an M3 with 24 GB of unified memory and an M4 with 32 GB. Both handle the full setup. For local models, the memory rules from [our model tests](/blog/running-openclaw-with-local-models) apply on top: 24 GB to experiment, 32 GB for everyday use. ## How to install OpenClaw on your Mac? ### Option 1: the official installer 1. **Run the official installer.** ```bash curl -fsSL https://openclaw.ai/install.sh | bash ``` It detects your system, installs Node when missing, and finishes by launching onboarding. Prefer your own package manager? The equivalent is: ```bash npm install -g openclaw@latest --allow-scripts=openclaw ``` On npm 11.15 or earlier, drop the `--allow-scripts` flag. The [official install page](https://docs.openclaw.ai/install) lists the pnpm and bun variants. 2. **Run the onboarding wizard.** If the installer didn't drop you into it: ```bash openclaw onboard --install-daemon ``` The wizard picks where the Gateway runs (local is what you want here), connects your AI provider (it detects credentials you already have, like a Claude Code login or a running Ollama, and offers OAuth sign-ins for several providers), and requests the macOS permissions it needs. The `--install-daemon` flag registers the Gateway as a launchd service, which is what keeps it running later. 3. **Check the Gateway is alive.** ```bash openclaw gateway status ``` Healthy looks like `Runtime: running` with a passing connectivity probe. `openclaw status` summarizes the whole install, and `openclaw logs --follow` streams what the Gateway is doing. When something looks off, `openclaw doctor` diagnoses it, and `openclaw doctor --fix` repairs the common cases. 4. **Say hello.** ```bash openclaw dashboard ``` This opens the built-in web chat through a single-use pairing link (the Gateway listens on `127.0.0.1:18789`). Send a first message. To chat from a messaging app instead, Telegram is the fastest channel to wire up (a bot token, nothing to install); the [channels docs](https://docs.openclaw.ai/channels) cover Telegram, WhatsApp, and the rest. ### Option 2: Ollama Ollama's [OpenClaw integration](https://docs.ollama.com/integrations/openclaw) bundles the whole setup, agent and local model together, into a single command. 1. **Install Ollama.** Download the application from [ollama.com](https://ollama.com) and open it. 2. **Launch OpenClaw.** Open a terminal and run: ```bash ollama launch openclaw ``` It installs OpenClaw and wires Ollama in as its provider, so your assistant thinks with a model running on your own Mac. 3. **Read the security notice and accept it.** On first launch Ollama walks you through the risks and asks for an explicit acknowledgement. Like the manual path, this installs OpenClaw directly on your Mac, running with your user's permissions, so everything in [where everything lives](#where-everything-lives) applies here too. The model choice matters more than the command: [we tested six local models](/blog/running-openclaw-with-local-models) to see which ones can actually drive OpenClaw, and the [Qwen 3.8 27B tutorial](/docs/tutorials/use-qwen-3-8-27b-with-openclaw) walks a complete local setup, including this route. ### Option 3: install with HolaClaw (native Mac app) [HolaClaw](/download) is our Mac app. It installs OpenClaw inside an isolated virtual machine with secure defaults, connects your provider (a cloud key or a managed local model), and handles updates and backups for you. No terminal, no account, about five minutes from download to first conversation. 1. **Download and open.** Get it from the [download page](/download) and drag HolaClaw to Applications. 2. **Click Get Started.** It begins the Create Assistant flow: pick a starting personality or shape your own. 3. **Pick your AI.** In the model provider step, bring an API key from your cloud provider, or choose a model from the local catalog and let HolaClaw manage it. 4. **Start talking.** Behind the scenes, HolaClaw creates the isolated VM, installs OpenClaw inside it, and applies the secure defaults. Your assistant is live in the app, and you can connect Telegram whenever you want it on your phone. **How is HolaClaw different from the other options?** - **Isolation.** OpenClaw runs inside a virtual machine, so the agent never touches your files directly. The permissions concern from [where everything lives](#where-everything-lives) doesn't apply. - **Day-2 handled.** Updates and backups are built in. Nothing in the update routine below is yours to remember. - **Fewer knobs.** The trade-off: the manual install exposes the whole config surface (channels, plugins, every setting), while HolaClaw curates it to keep the setup safe and simple. [Download HolaClaw](/download) to try it. ## Where everything lives Everything OpenClaw knows sits in `~/.openclaw`. The file you'll actually touch is `openclaw.json`, the Gateway's config: it's JSON5, so comments are allowed, and it's validated strictly. A malformed or unknown key stops the Gateway from starting, and `openclaw doctor --fix` repairs it. Most edits hot-reload without a restart. One thing to be clear about, because it shapes everything else on this page: **installed this way, OpenClaw runs as you, with your user's permissions.** Whatever your account can read, the agent can read. The official [security docs](https://docs.openclaw.ai/gateway/security) are direct about the consequence: any untrusted content the agent reads (web pages, emails, attachments, pasted logs) can carry adversarial instructions. Keep the Gateway on loopback (never expose it unauthenticated on `0.0.0.0`), leave the messaging pairing policies on their defaults, and consider real isolation if your agent will browse the web for you. That risk is the reason HolaClaw runs OpenClaw inside a virtual machine instead. ## How to keep it running? If you onboarded with `--install-daemon`, this is already done: the Gateway is registered as a launchd agent (`ai.openclaw.gateway`) that starts when you log in and restarts if it crashes. If you skipped it, `openclaw gateway install` registers the same service. The commands worth knowing: ```bash openclaw gateway status # is it running, and is it reachable? openclaw gateway restart # after config changes that don't hot-reload openclaw gateway stop --disable # stop it, and keep launchd from respawning it ``` Closing the terminal doesn't stop the Gateway; launchd keeps it alive in the background until you disable it. ### On a dedicated Mac (like a mac mini) OpenClaw fits the closet-server pattern well: install it on a spare, always-on Mac, and talk to it from every other device through a channel like Telegram. The install is identical to the steps above. Two extra things matter: keep the machine from sleeping (System Settings, Energy), and reach the dashboard from your other devices over [Tailscale](https://tailscale.com) rather than binding the Gateway to the local network, which is also what OpenClaw's security docs recommend. ## How to update without breaking your setup? Updates are the sharp edge of running OpenClaw yourself. The project moves fast (stable releases every week or three, with beta and extended-stable channels around them), and an update that lands overnight can change behavior you relied on. The routine that keeps it boring: 1. **Back up first.** ```bash openclaw backup create --output ~/openclaw-backup --verify ``` 2. **Update.** ```bash openclaw update ``` It detects how you installed (npm, pnpm, bun, or git), fetches the latest release, runs `openclaw doctor`, and restarts the Gateway. 3. **Verify.** `openclaw health`, then a test message. If an update misbehaves, roll back by pinning a known-good version: `openclaw update --tag ` (add `--dry-run` first to preview). And if you'd rather trade freshness for stability, `openclaw update --channel extended-stable` switches you to the slow lane, which never applies updates automatically. ## If something doesn't work - **The Gateway won't start and the logs mention the port is in use.** Something else, often an older Gateway instance, is already listening on 18789. Stop it, or change `gateway.port` in `openclaw.json`. In HolaClaw, errors will be surfaced in the GUI although you can always [connect to your virtual machine](/docs/advanced/access-the-virtual-machine). - **The service is installed but nothing is running.** Check `openclaw gateway status`. A config whose `gateway.mode` points at a remote Gateway means nothing starts locally; set it back to local mode. - **The Gateway refuses to start after you changed where it listens.** Binding beyond loopback without auth configured is rejected by design. Configure a gateway auth token, or go back to `127.0.0.1` and use Tailscale for remote access. - **The Gateway refuses to start after a config edit.** That's the strict validation. `openclaw doctor --fix` repairs it, and the error names the offending key. - **Onboarding can't detect the Gateway.** Run `openclaw gateway status` in a second terminal, and `openclaw logs --follow` will usually name the reason it isn't up. - **A fresh install is missing pieces, like the dashboard.** Broken package releases happen. Pin the previous release with `openclaw update --tag `, or reinstall through the official script. ## Questions people ask **Is OpenClaw free?** Yes, it's free and open source. What can cost money is the model behind it: cloud providers bill per token, while a local model runs at no cost beyond your own hardware. **Do I need an API key?** No. A cloud provider needs a key or an OAuth sign-in during onboarding, but OpenClaw also runs against local models served by Ollama, with no key and no account. [Local models](/docs/ai-providers/local-models) covers what that takes. **Does it work offline?** With a local model, the AI itself runs entirely on your Mac. The channels decide the rest: the local web chat works offline, while Telegram and WhatsApp need the network by nature. **Is OpenClaw safe to run on my Mac?** Installed manually, it runs with your user's permissions and can touch whatever your account can. The project's own security docs are upfront about that, and the defaults are sane: Gateway on loopback, messaging locked behind pairing. However, it still gives the agent all the permissions you already have. The stronger answer is isolation, running the agent inside a container or a virtual machine. That's the design HolaClaw ships by default. **Can I uninstall it cleanly?** Yes. `openclaw uninstall` removes the install, and `openclaw uninstall --dry-run --all` previews everything it would touch first. The [official uninstall page](https://docs.openclaw.ai/install/uninstall) also documents the manual cleanup, including the `~/.openclaw` state directory and the launchd service. **Does it need a powerful Mac?** It depends on where the model runs. With a cloud provider, the heavy lifting happens on their servers and the Gateway is a light Node service. With local models, memory is the limit: [our tests](/blog/running-openclaw-with-local-models) put the practical floor at 24 GB. ## Related - [Which local models can run OpenClaw on your Mac? We tested 6](/blog/running-openclaw-with-local-models): what runs, what breaks, and the RAM you need. - [Automatic backups for OpenClaw](/blog/automatic-backups-for-openclaw): why we back up before every update, with the failure stories. - [Choosing an AI provider](/docs/ai-providers/choosing-a-provider): cloud against local, and how to pick. - [Local models](/docs/ai-providers/local-models): running your assistant with no account, no API key, no cloud. Hit an install problem this page doesn't cover? Tell us in our [Discord](https://discord.gg/FbxAbS5sGQ) and we'll add it to the guide.