OpenClaw: what it is and how to start
Updated on
OpenClaw is an open-source AI assistant that runs on your own computer and meets you in the chat apps you already use. This guide covers what the pieces are, what you need before the first command, how the official install starts, and the security defaults to read before anyone else can message it.
Status: Not tested yet. We have not installed OpenClaw or sent it a message, so the steps and product facts below come from the OpenClaw docs, the project README, the OpenClaw website, the GitHub releases page, and the Grok Bot docs. Sources checked on 2026-10-31.
What OpenClaw is
OpenClaw is an open-source AI assistant that runs on your own computer and meets you in the chat apps you already use. The README names Discord, iMessage, Slack, Teams, Telegram, WhatsApp, and 20+ more, plus native apps for macOS, iOS, Android, Windows, and Linux. One Gateway can be a personal assistant on a laptop or a shared team deployment. Configuration is the only difference.
State, memory, and credentials live on your hardware. Claude, Codex, and local models are plugins you can swap. Prompts go to the model provider and the chat platforms you configure, plus any diagnostics export you turn on. By default OpenClaw phones home for a daily version check. Anonymous feature statistics are opt-in. update.checkOnStart: false disables both.
Peter Steinberger and the community built it. Today the OpenClaw Foundation, an independent US 501(c)(3), stewards the project, employs the core team and signs releases. The license is MIT, with no enterprise edition, no paid tier, no hosted service and no token. Donors include the University of Michigan, OpenAI, Amazon and Red Hat; the website says OpenAI is a donor and Peter Steinberger works there, but OpenClaw is not an OpenAI product and donors don't direct the project.
On 2026-10-31 the latest GitHub release was v2026.10.9, published 2026-10-29.
If you have seen the names Clawdbot or Moltbot, that is the same project. The lore page says it started as Warelay, a WhatsApp gateway, then became Clawdbot. After Anthropic asked for a name change in January 2026 (trademark), it was renamed Moltbot on January 27, 2026 and OpenClaw on January 30, 2026.
How it fits together
| Piece | What it does |
|---|---|
| Gateway | Local control plane for sessions, tools, events, and channels. A regular host install binds it to loopback. |
| Control UI, CLI, TUI | Clients of the Gateway. openclaw dashboard opens the web UI. openclaw opens the TUI. |
| Channels | Chat apps through the Gateway. Text is supported. Media and reactions vary. |
| Apps and nodes | Voice, Canvas, camera, screen, and device-local actions. A desktop app can start a local Gateway or connect to a remote one. |
| Model providers | Hosted and local models. Adding provider auth does not change the primary model. |
| Skills, plugins, ClawHub | SKILL.md folders, plugins on a small core, and the ClawHub registry. |
A few details worth knowing before you install:
- Channels. Bundled channels ship with the install; others are official plugins you add with
openclaw channels add. Several can run at once. The channels page has the full list, including Discord, iMessage, Matrix, Microsoft Teams, Signal, Slack, Telegram and WhatsApp. - Apps. There is a macOS menu bar app and a Windows Hub app, and the desktop apps can start a local Gateway or connect to a remote one.
- Models. The model-providers page lists OpenAI, Anthropic, Google Gemini, xAI, OpenRouter and NVIDIA, plus local servers such as Ollama, LM Studio, llama.cpp and vLLM. Onboarding can also reuse an existing Claude Code or Codex CLI login.
- Skills. A skill is a folder with a
SKILL.mdfile, following the AgentSkills spec. ClawHub is the public registry:openclaw skills install @owner/<slug>. The docs say to treat third-party skills as untrusted code, and note that native plugins run in-process without a sandbox.
What you need
- A computer running macOS, Linux or Windows (Windows Hub, PowerShell, or a Gateway in WSL2).
- Node.js 24.16.0+ or 26.1.0+, with Node 26 recommended. The installer provisions Node if it's missing, so you only need to care if you install with npm yourself.
- A model: an API key from a provider, or an existing Claude Code or Codex CLI login. That provider bills you separately.
- Optional: a chat app account and bot token, such as a Telegram bot, if you want to message it from your phone.
Install and first run
Read Security basics before other people can message the agent.
Install
On macOS, Linux, or WSL2, run curl -fsSL https://openclaw.ai/install.sh | bash. On Windows PowerShell, run iwr -useb https://openclaw.ai/install.ps1 | iex. The script installs Node if needed, installs OpenClaw, and launches onboarding.
curl -fsSL https://openclaw.ai/install.sh | bash
iwr -useb https://openclaw.ai/install.ps1 | iex
The install page calls this the fastest path. If you already manage Node, you can install the npm package instead (on npm 11.15 and earlier, drop the --allow-scripts flag), then run openclaw onboard --install-daemon.
npm install -g openclaw@latest --allow-scripts=openclaw
Run onboarding
Choose Quick start to reuse a detected Claude Code or Codex login or an API key, or Custom setup for the full flow. Then run openclaw gateway install so the Gateway keeps running after you leave the terminal.
Quick start looks for a Claude Code or Codex CLI login or an API key, checks it with a real request, saves the config and opens the dashboard. Custom setup walks through every option, and you can come back later with openclaw configure.
Quick start keeps the Gateway in the foreground. Press Ctrl+C, then:
openclaw gateway install
That installs the Gateway as a background service: a LaunchAgent on macOS, a systemd user unit on Linux and WSL2, or a Scheduled Task on Windows.
Check the Gateway and open the dashboard
Run openclaw gateway status and confirm the Gateway is listening on port 18789. Run openclaw dashboard to open the Control UI, then send a message.
openclaw gateway status
openclaw dashboard
If the reply comes back, onboarding has verified model access, created the workspace, and configured the Gateway.
If the openclaw command isn't found, the docs say it is almost always a PATH problem. openclaw doctor runs health checks and repairs config.
Connect a chat channel
Start with Telegram. Run openclaw channels add --channel telegram --token followed by your bot token, or run openclaw channels add with no flags and pick a channel from the list.
openclaw channels add --channel telegram --token <bot-token>
Telegram needs only a bot token, no plugin install. Check the connection with openclaw channels status --probe. WhatsApp pairs with a QR code instead.
Security basics
Treat inbound messages as untrusted input. Most channels answer an unknown DM with a pairing code instead of processing it:
openclaw pairing approve <channel> <code>
Tools run on the host for the main session unless you configure sandboxing. Read the security guide, the exposure runbook, and the sandboxing guide before other people can message the agent or the Gateway is reachable off the machine.
A normal install binds the Gateway to loopback, so it is not reachable from other machines. Group chats are allowlisted, usually behind a mention. openclaw security audit reports settings that have drifted from the defaults. Docker images are the exception: they default to an exposed bind, which the docs say to pair with auth.
Sandboxing and exec approvals are off by default, because OpenClaw assumes one trusted operator. A sandbox moves tool execution into an isolated environment; the docs say it is not a perfect boundary but limits file and process access. One Gateway is not meant to separate people who don't trust each other: for that, the docs say to run a separate Gateway with separate credentials, ideally under a separate OS user or host.
What it costs
The Foundation is funded by donations and sells nothing: no paid tier, no hosted service, no token. You still bring an API key or a Claude Code or Codex login, and that provider bills you. Prices are left out on purpose.
OpenClaw vs a hosted agent like Grok Bot
OpenClaw and Grok Bot are both personal agents you message, but one runs on your hardware and the other is a hosted service. We have not run either for this page; the table is from each one's docs.
| Topic | OpenClaw | Grok Bot |
|---|---|---|
| Where work runs | Your computer. State, memory, and credentials stay there. | A persistent cloud computer in Cursor's cloud: browser, filesystem, and terminal. |
| How you reach it | Control UI, CLI, TUI, chat channels, native apps. | Message the Bot on macOS, Windows, Linux, iPhone, or Android. |
| Laptop closed | A background service can run on that machine. The docs do not say a closed laptop keeps working. | Closing the app, laptop, or phone does not stop a background turn or a routine. |
| Skills | SKILL.md folder, including ClawHub installs. | Walk a path once, save it as a skill, rerun it on a schedule. |
| Product bill | No paid tier. No hosted service. The model provider bills you. | X Premium, Super, and Ultra; paid legacy Cursor plans and Cursor Teams; or a linked SuperGrok, SuperGrok Plus, SuperGrok Heavy, X Premium+, or Grok Business seat. |
| Sharing | One gateway, one trust boundary. | Your Bots share one computer and its files. Between users, isolation is strict. |
The trade-off is control versus upkeep. OpenClaw keeps your state and credentials on your own machine and lets you choose any model, but you install updates, secure the Gateway and keep the machine running. Grok Bot handles the computer for you and connects to apps through plugins. Prices are left out on purpose; the plan list is on the Grok Bot plans page. See Grok Bot skills and Grok Bot plugins.
Limits of this guide
- We have not installed OpenClaw or sent it a message. Everything above comes from the README, the website and docs.openclaw.ai, checked on 2026-10-31. If we test it, the run will be logged here with the date, version and output, following how we test.
- OpenClaw ships often, so commands and defaults can change after v2026.10.9. This guide skips VPS, Docker and Nix installs, sandbox settings, and how to create a Telegram bot token.
- If you are moving off OpenClaw, the Hermes Agent README says Hermes can import OpenClaw settings, memories and skills (and API keys, if you ask). We have not tried that import.
FAQ
What is OpenClaw?
OpenClaw is an open-source AI assistant that runs on your own computer and talks to you in chat apps you already use. A local Gateway is the control plane for sessions, tools, and channel connections. The OpenClaw Foundation, an independent 501(c)(3), stewards the project. Peter Steinberger and the community built it. The license is MIT.
Is OpenClaw free?
OpenClaw has no paid tier, no hosted service, and no token. The Foundation says it sells nothing and is funded by donations. You still bring a model, through a provider API key or an existing Claude Code or Codex login, and that provider's plan is separate. Prices are left out on purpose.
Is OpenClaw owned by OpenAI?
No. The OpenClaw website says OpenAI is a donor and that Peter Steinberger works there, and that OpenClaw is not an OpenAI product. Donors do not own, control, or direct the project. The OpenClaw Foundation stewards it.
How do I install OpenClaw?
On macOS, Linux, or WSL2, the docs recommend curl -fsSL https://openclaw.ai/install.sh | bash. On Windows PowerShell, they recommend iwr -useb https://openclaw.ai/install.ps1 | iex. The script installs Node if it is missing and starts onboarding. You need Node 24.16.0 or newer on the 24.x line, or Node 26.1.0 or newer. Node 26 is recommended, and newer patch releases are recommended for security fixes.
Is it safe to let people message an OpenClaw bot?
Treat inbound messages as untrusted input. On a normal host install the Gateway binds to loopback, and most chat channels answer an unknown DM with a pairing code instead of processing it. Approve a pairing with openclaw pairing approve, plus the channel name and the code. Tools run on the host for the main session unless you turn sandboxing on, and sandboxing is off by default. Read the security guide before you expose the Gateway or connect other people. Container images are an exception: they default to an exposed bind.
How is OpenClaw different from Grok Bot?
OpenClaw is software you run. It has no paid tier. State, memory, and credentials live on your hardware, and your prompts go to the model provider you configure. Grok Bot is a hosted product. Each Bot works on a persistent cloud computer in Cursor's cloud, and access comes with certain X, Cursor, SuperGrok, or Grok Business plans. Grok Bot keeps working when your laptop is closed. The OpenClaw docs describe a background Gateway service on the machine where you installed it, and they do not say a closed laptop keeps working.
Related on bests.one
- What is Grok Bot? — a hosted agent on a cloud computer.
- Grok Bot skills — saved instructions inside Grok Bot, not SKILL.md folders.
- Grok Bot plugins — how a hosted Bot connects to apps.
- How to create a Grok Bot
- How we test
Sources: OpenClaw README (GitHub), OpenClaw releases (GitHub), OpenClaw website, Getting started (OpenClaw docs), Why OpenClaw (OpenClaw docs), The lore of OpenClaw (OpenClaw docs), Install (OpenClaw docs), Channels (OpenClaw docs), Skills (OpenClaw docs), Model providers (OpenClaw docs), Gateway security (OpenClaw docs), Sandboxing (OpenClaw docs), FAQ (OpenClaw docs), Grok Bot overview (docs), Grok Bot plans and billing, Hermes Agent README (GitHub). Checked on 2026-10-31.