Self-hosting OpenClaw
OpenClaw is an open-source, MIT-licensed personal agent gateway: a single process that links your chat apps to the model providers you pick and runs tools for you. This section shows how to install it on a laptop, a VPS, Docker or a Raspberry Pi, connect models and Telegram, reach it remotely and lock it down.
- Latest release
2026.9.7· September 30, 2026- Extended stable
2026.8.33- License
- MIT · openclaw/openclaw · 391,000 GitHub stars
- Runtime
- Node.js 24.16+ / 26.1+
- Minimum host
- 1 vCPU, 1 GB RAM, 500 MB disk
- Default port
127.0.0.1:18789- Config file
~/.openclaw/openclaw.json
Quick install
The installer script sets up Node and the openclaw CLI, then starts onboarding. It provisions Node 26 on macOS and Node 24 LTS on Linux; on native Windows, use the PowerShell installer.
curl -fsSL https://openclaw.ai/install.sh | bashiwr -useb https://openclaw.ai/install.ps1 | iexNever publish the gateway port to the internet: OpenClaw binds to loopback and requires a token by default, so leave both in place. Internet scans during 2026 have counted exposed OpenClaw gateways in the tens of thousands.
Step-by-step guides
Install OpenClaw
Install the openclaw CLI with the official script, the PowerShell installer or npm, then let onboarding create the config, the gateway token and the background service.
Run OpenClaw on a VPS
Put OpenClaw on a small Linux server so the agent stays online when your laptop is off, without opening the gateway to the internet.
Run OpenClaw in Docker
Run the gateway and the CLI as Docker Compose services, built from source or pulled as the prebuilt ghcr.io/openclaw/openclaw image.
Run OpenClaw on a Raspberry Pi
A Raspberry Pi works as a quiet, always-on OpenClaw host, as long as the models run somewhere else.
Connect models to OpenClaw
Give OpenClaw an API key for a cloud provider and choose the default model, or point it at a local server such as LM Studio or vLLM.
Connect OpenClaw to Telegram
Telegram ships in OpenClaw core. Create a bot, give the gateway its token and approve your own account with a pairing code.
Harden your OpenClaw gateway
Apply the hardened baseline from the OpenClaw docs: loopback and token auth, a sandbox, strict exec approvals and tight control over who can message the agent.
Remote access to OpenClaw
Reach the dashboard of a gateway on another machine with an SSH tunnel or Tailscale, without changing the loopback bind.
Hardened baseline
These values come from the hardened baseline in the OpenClaw docs. Start from them, relax a setting only when a workflow needs it, and run the security audit again whenever you change one.
| Setting | Value | Why it matters |
|---|---|---|
gateway.bind | loopback | Only processes on the same machine can reach the gateway. Use an SSH tunnel or Tailscale for remote access. |
gateway auth | token | Every client must present the token. Use 24 characters or more; the audit warns on shorter tokens. |
session.dmScope | per-channel-peer | Each sender on each channel gets a separate session, so no one sees another person's context. |
tools.profile | messaging | Starts the agent from a narrow, messaging-focused tool set instead of every available tool. |
fs.workspaceOnly | true | Keeps file tools inside the agent workspace, away from SSH keys, credentials and the rest of the disk. |
exec.security | deny | Blocks shell commands outright. On gateway hosts the default is full, which allows any command. |
exec.ask | always | If you later relax exec.security, every command waits for approval; askFallback denies when nobody answers. |
elevated.enabled | false | Turns off the /elevated chat command, so no sender can switch the agent into elevated mode. |
dmPolicy | pairing | New senders receive a code that you must approve. Codes expire after an hour; at most three wait. |
groups.requireMention | true | In group chats the agent answers only when mentioned, not every message anyone posts. |
Chat channels
Telegram, WebChat, A2A and Reef ship in core; the others are official plugins. Telegram uses long polling by default, so it needs no public webhook URL.
Name history
Peter Steinberger started the project in November 2025 under the name Clawdbot. It became Moltbot on 27 January 2026 after a trademark request from Anthropic, and OpenClaw two days later, on 29 January 2026. The OpenClaw Foundation, a US 501(c)(3) non-profit, was announced on 8 July 2026.
OpenClaw questions
Is OpenClaw free to self-host?
Yes. OpenClaw is MIT-licensed and its source is on GitHub at openclaw/openclaw. You pay for the machine it runs on and for any model API calls, unless you serve models locally.
What are the minimum requirements for OpenClaw?
The docs list 1 vCPU, 1 GB RAM and about 500 MB of disk, with Node.js 24.16+ or 26.1+. They recommend 1–2 vCPU and 2 GB RAM or more. These figures assume cloud model APIs; running models locally needs far more memory.
Which port does the OpenClaw gateway use?
Port 18789, bound to 127.0.0.1 by default. One process serves WebSocket, HTTP and the dashboard on it. You can change it with gateway.port, the --port flag or OPENCLAW_GATEWAY_PORT, but do not open it to the internet.
Does OpenClaw run on Windows?
Yes. The install page lists the native Windows Hub app, the PowerShell installer and WSL2. On Windows the background service is a Scheduled Task. Parts of the FAQ still mention only WSL2.
Is OpenClaw safe to run on my machine?
It can run commands and touch files on its host, so the settings matter. Gateway auth is required and the bind is loopback by default, but sandboxing is off and exec.security is full on gateway hosts. Apply the hardened baseline and run openclaw security audit --deep.
Where does OpenClaw keep its configuration?
The config file is ~/.openclaw/openclaw.json and the agent workspace is ~/.openclaw/workspace. Include both in your backups and copy them when you move to a new host.