How to Install OpenClaw 2.0 (Or Skip the Setup Entirely)

OpenClaw shipped its biggest release ever at the end of August 2026. Version 2026.8.1, which the project calls OpenClaw 2.0, folds in more than 16,000 pull requests from 933 contributors, roughly half of everything ever merged into the project. The headline change is the one most people searching for how to install OpenClaw care about: the install experience was rebuilt. The old "clone the repo and pray the Docker script works" flow is gone. Docker is now optional, a one-line installer handles Node for you, and the onboarding wizard finds AI access you already have on the machine instead of demanding a fresh API key.

That is real progress, and this guide walks through it step by step. It also covers what did not change: you still need a computer that never sleeps, you still pay for model access, you still own the security of an agent that can run shell commands, and your agent still cannot pick up a phone. If you only want an AI agent that does things, there is a shorter path at the end.

Quick Answer: How to Install OpenClaw in 2026

• Fastest path (macOS, Linux, WSL2): curl -fsSL https://openclaw.ai/install.sh | bash. The script installs a compatible Node.js if needed, then starts onboarding.
• Windows: open PowerShell and run iwr -useb https://openclaw.ai/install.ps1 | iex, or use the Windows desktop installer from the project's releases page.
• Package manager: npm install -g openclaw@latest --allow-scripts=openclaw, then openclaw onboard --install-daemon.
• Docker: still supported, but it is now one deployment option among many (Kubernetes, Fly.io, Render, DigitalOcean), not the default.
• Requirements: Node.js 22.22.3+, 24.15+, or 25.9+ (Node 26 recommended), a machine that stays on, and model access (a ChatGPT or Claude CLI sign-in, an API key, or a local model through Ollama or LM Studio).
• Time: 15 to 30 minutes for a clean install if everything detects correctly; longer if you are upgrading from an older version or want sandboxing.

What Changed in OpenClaw 2.0 for Installation

If you read an OpenClaw install guide from earlier this year (including the previous version of this one), most of the commands in it are now stale. Here is what actually changed.

Docker is no longer required. The February-era flow was git clone, then ./docker-setup.sh, then a manual docker compose fallback when the script failed. OpenClaw now installs as a normal Node.js CLI with a managed background service. Docker remains available for containerized and cloud deployments, but a personal install on a Mac Mini or a VPS no longer needs it. This is the single most-searched OpenClaw install question, and the answer is finally yes: you can install OpenClaw without Docker.

The installer handles Node.js. The install script detects your operating system and installs a supported Node version if you do not have one. You no longer need to sort out nvm or Homebrew first.

Onboarding detects the AI access you already have. The guided model setup looks for existing Codex, ChatGPT, or Claude command-line sign-ins, saved API keys, configured model routes, and local models running through Ollama or LM Studio. It then fires a real completion request against your chosen model and only saves the configuration once the model actually answers. The old wizard would happily save a broken key and let you discover it later.

Setup moves into the conversation. The foundation's stated goal for 2.0 is "letting people get to a first conversation faster and finish setting up their Claw by talking to it." Non-essential configuration, like messaging channels and skills, is deferred until after you are already chatting with the agent in the browser.

The browser app is the main interface. The control UI is no longer a settings panel you visit once to paste a token. It is a ChatGPT-style workspace with a conversation sidebar, a workspace file editor, a git-backed Changes panel, and a web terminal. Startup was cut from roughly 1.6 seconds to 575 milliseconds in the project's own benchmarks.

A managed service keeps it running. openclaw onboard --install-daemon registers a LaunchAgent on macOS, a systemd unit on Linux and WSL2, or a Scheduled Task on Windows. That closes the biggest complaint from the Docker era: the agent silently dying when the machine rebooted.

Step 1: Check the Prerequisites

You need the following before you start:

• A computer that stays on. OpenClaw runs as a background gateway. If the machine sleeps, your agent is offline. Most people use a dedicated Mac Mini, a small Linux box, or a $10 to $40 per month VPS. The project's security guidance recommends a dedicated host with full-disk encryption, or at minimum a separate operating-system account.
• Node.js 22.22.3+, 24.15+, or 25.9+. Node 26 is the recommended version. The install script can handle this for you; the npm route assumes you already have it.
• Model access. One of: an existing ChatGPT or Claude CLI sign-in, an API key from OpenAI, Anthropic, Google, or OpenRouter, or a local model through Ollama or LM Studio. The 2.0 defaults are GPT-5.6 for new cloud setups and Gemma 4 with a 64K context window for local runs, gated on how much RAM you have.
• Terminal comfort. Fewer commands than before, but you will still run a shell, read a log, and occasionally run openclaw doctor.

If you are on Windows, install WSL2 first or use the native PowerShell installer. WSL2 gets you the same experience as Linux and is what most Windows users end up on.

Step 2: Install OpenClaw With the Script (Recommended)

Open a terminal and run:

curl -fsSL https://openclaw.ai/install.sh | bash

The script detects your OS, installs Node if required, installs the openclaw CLI globally, and launches onboarding. If you want to install now and configure later (useful when scripting a server build, or when a previous onboarding half-completed), add the no-onboard flag:

curl -fsSL https://openclaw.ai/install.sh | bash -s -- --no-onboard

On Windows PowerShell:

iwr -useb https://openclaw.ai/install.ps1 | iex

If you would rather keep everything under ~/.openclaw without a system-wide Node, there is a local-prefix variant at https://openclaw.ai/install-cli.sh. It is the cleanest option on a shared machine.

Step 3: Install OpenClaw Without Docker Using npm, pnpm, or Bun

If you already manage Node yourself, skip the script and use your package manager. All three routes end with the same onboarding command.

npm (11.15 or newer):

npm install -g openclaw@latest --allow-scripts=openclaw
openclaw onboard --install-daemon

pnpm:

pnpm add -g --allow-build=openclaw openclaw@latest
openclaw onboard --install-daemon

Bun (1.4 or newer):

bun add -g --trust openclaw@latest
bun run --bun openclaw onboard --install-daemon --daemon-runtime bun

The --allow-scripts and --allow-build flags are new hardening in package managers, not something OpenClaw invented. OpenClaw runs a post-install step, and modern npm and pnpm refuse to run those silently, so you must opt in by name.

Step 4: Run Onboarding and Verify

Whichever route you took, onboarding does the following:

1. Scans for existing model access and offers what it finds. Pick a Claude or ChatGPT sign-in, paste an API key, or point at a local Ollama model. 2. Sends a test request and waits for a real answer before saving anything. 3. Provisions a local Gateway, bound to loopback (127.0.0.1) by default so nothing is exposed to your network. 4. Installs the managed startup service if you passed --install-daemon. 5. Opens the browser app, where you finish setup by talking to the agent.

Then confirm the pieces are alive:

openclaw --version
openclaw doctor
openclaw gateway status

openclaw doctor is the command you will use most. It checks configuration, flags stale settings, and with --fix migrates old formats for you. If you see "missing config, run openclaw setup or set gateway.mode=local," that means onboarding never completed; rerun openclaw onboard.

Step 5: Upgrading From an Older OpenClaw (Read This First)

The 2.0 upgrade is where most of the early pain reports come from. Reception has been mixed: plenty of clean upgrades, but also broken gateways, lost automations, and authentication failures on machines that skipped a step. Do these in order.

Back up ~/.openclaw before anything else. The release notes say so explicitly, and the reason is the next point.

Sessions moved from files to SQLite. The migration archives your old file-based transcripts. If you roll back to a pre-2.0 release, sessions created after the migration will not appear until you restore the archived artifacts. Treat the upgrade as one-way unless you have a backup.

Run the doctor after updating:

openclaw update --channel stable
openclaw doctor --fix

The --fix pass handles two breaking changes. The bundled OpenProse plugin and its /prose command were removed, so the doctor cleans that configuration and points you at the Agent Skills migration path (your .prose files are kept). And model references under codex/* and openai-codex/* are rewritten to the new openai/* routes.

"Gateway service already installed" errors. If the installer complains that a gateway service exists and you want a clean start, stop the old service, remove the stale unit or LaunchAgent, and rerun openclaw onboard --install-daemon. Reinstalling the CLI on top of a running old gateway is the most common way to end up with two half-working copies.

Check your automations. Scheduled jobs are the thing users most often report losing across the upgrade. Open the browser app, confirm each one is still listed, and re-approve any that show as needing fresh permission. In 2.0, an automation's approval is tied to the exact operation it performs; if the operation changed, it needs a new approval.

Step 6: Security Settings You Should Change on Day One

OpenClaw 2.0 added real security tooling. It also shipped several of those tools switched off, which is why The Register headlined its coverage "OpenClaw 2.0 pours glitter on slow-burning security dumpster fire." The truth is in between: the tooling is genuinely better, and the defaults still favor getting to a first conversation over locking things down. Change these right after install.

• Run the audit. openclaw security audit checks inbound access, the tool "blast radius," network exposure, browser-control exposure, and plugin allowlists. Do it before connecting any messaging channel.
• Turn on code sandboxing. Isolation for untrusted code exists in 2.0 but is off by default. Enable it if the agent will run anything it did not write itself.
• Know what the secret store does and does not do. The new protected-credentials feature keeps API keys out of chat and out of the model's context, which is a big improvement. But per the release notes, "Secret Store values are not encrypted at rest and depend on the filesystem permissions of OpenClaw's state directory." Full-disk encryption on the host is what protects them.
• Do not treat shared sessions as isolation. The multiplayer feature lets teammates view, suggest, or write in a live session. The documentation states that these controls "are not tenant isolation and not a security boundary." One Gateway is one trust boundary. Fine for a single person or a small team that already trusts each other; not fine for serving customers.
• Review every plugin. External plugins now go through a capability review that shows source, version, and artifacts before install. Read it. Malicious skills were a real attack vector in the February security incidents, and the review screen only helps if you look at it.
• Expect prompt injection to land sometimes. The project's own testing across 272,000 injection attempts put success rates between 0.5% and 8.5% depending on the model, and it warns that adaptive human attackers still exceed 80% against state-of-the-art defenses. Keep human approval on for anything that spends money, sends messages, or deletes files.

The project's own summary of who should run it is refreshingly blunt: yes for single-operator and single-team deployments, no for multi-tenant products.

Step 7: Connect Messaging Channels

Channel setup now happens inside the browser app after onboarding rather than in the CLI wizard, but the requirements are the same as before. WhatsApp pairs by QR code. Telegram needs a bot token from BotFather. Discord needs an application in the developer portal. Slack, Signal, and iMessage are also supported. Connect one at a time and run the security audit again afterward, because each channel is a new inbound path to an agent with shell access.

Common Problems After Installing OpenClaw 2.0

• Onboarding says no model access found. Sign in to the Claude or ChatGPT CLI first, or export an API key, then rerun openclaw onboard. Detection only finds what is already on the machine.
• Node version errors. The npm route does not install Node for you. Use the install script, or upgrade to Node 26 and retry.
• Gateway status shows stopped after reboot. You skipped --install-daemon. Rerun onboarding with the flag so the LaunchAgent, systemd unit, or Scheduled Task gets created.
• Old sessions missing. You rolled back to a pre-2.0 build. Either move forward again or restore the archived legacy files.
• High memory use. Local models are the usual cause. The 2.0 local default is RAM-gated, but a 64K context window on a small box will still swap. Switch to a hosted model or a smaller local one.
• Costs creeping up. Long-running background agents burn tokens while you sleep. Check usage in the browser app weekly; the project itself warns that background agents can consume "significant tokens, compute and external-service quotas."

What It Actually Costs to Run OpenClaw

The software is free under the MIT license. Everything around it is not:

• Hardware: a Mac Mini or small PC you dedicate to it, or a VPS at $10 to $40 per month.
• Model access: a ChatGPT or Claude subscription you already pay for, API usage that scales with how busy the agent is (commonly $15 to $100+ per month), or free local models that need a machine with real RAM.
• Electricity: $5 to $15 per month for an always-on home box.
• Your time: 2.0 cut a clean install to well under an hour, but upgrades, plugin reviews, security audits, and the occasional broken gateway are ongoing. Budget an hour or two a month.

Call it $30 to $150 per month before the agent has done a single useful thing, plus the maintenance. For a full breakdown of how self-hosted stacks compare with subscription AI assistants, see our AI assistant cost guide.

What OpenClaw Still Cannot Do

OpenClaw 2.0 is a serious upgrade for anyone who wants a self-hosted AI agent that browses the web, edits files, runs code, and lives in your chat apps. It is now much easier to install and much more pleasant to use. Six months after OpenAI's acqui-hire of its creator, the open-source project is healthier than most people predicted.

It still lives entirely in the digital world. Even fully configured, OpenClaw cannot:

• Make or receive real phone calls
• Press through an IVR menu ("press 2 for billing")
• Wait on hold with your insurance company or a clinic and hand you the outcome
• Notify you on your phone when a task finishes, unless you wire up a messaging channel yourself

For the tasks people most want to hand off, which are overwhelmingly phone tasks, a self-hosted agent gets you to the point where you still have to make the call.

The Alternative: Skip All of This

Assindo is a managed AI agent that does what OpenClaw does in the digital world (web search, scheduling, social posting, task tracking) plus the part no self-hosted agent covers: it makes real phone calls. It dials the business, navigates the phone menu, waits on hold, talks to the human, and sends you a summary and transcript. It also answers and screens your incoming calls.

There is nothing to install. Download the app on iOS or Android, or open it in a browser. Your agent runs on maintained cloud infrastructure with push notifications, support for 29 languages, and a choice of assistant personas. No Node versions, no gateway service, no doctor --fix, no security audit, no 2 AM debugging. You pay a flat monthly subscription and get an AI assistant that works the moment you sign in.

If you enjoy running your own stack, OpenClaw 2.0 is the best it has ever been, and this guide should get you there in under an hour. If you just want the calls made, the shortest install is no install. For a wider look at what managed and self-hosted agents can each do today, see our best AI agent apps in 2026 roundup.