Most "AI agent" products are a subscription and a chat window on a website you have to remember to open. OpenClaw takes a different shape: it lives inside WhatsApp, Telegram, or Discord - apps already open on your phone - and beyond conversation, it can read and write files and run shell commands on the server it's deployed to. Message it like you'd message a person, and it can actually go do the thing.
That capability is exactly why it belongs on infrastructure you control, not a third party's.
What it needs to run
DeployOS deploys OpenClaw as a single container (ghcr.io/openclaw/openclaw) with a persistent volume for its state and a generated gateway token for authentication. From the Templates page:
- Find OpenClaw under the AI category, pick a server, deploy.
- DeployOS generates an
OPENCLAW_GATEWAY_TOKENautomatically and stores it in the same encrypted secret store as your database passwords - you don't paste one in by hand, and it's never written to a log. - Connect the running instance to WhatsApp, Telegram, or Discord using that platform's normal bot-connection flow.
That's the whole setup. The part worth slowing down on is what comes next.
Treat the gateway token like a root credential
An agent that can run shell commands is, functionally, a remote shell with a chat interface in front of it. Anyone who has OpenClaw's gateway token can ask it to act on your server - read a file, run a script, anything you've allowed it to do. That's not a flaw in OpenClaw, it's the entire point of the product; it just means the token deserves the same handling as an SSH key, not the same handling as a random app password.
A few things worth doing before you connect it to anything:
- Don't share the messaging thread it's connected to with people you wouldn't hand a terminal to.
- If you're experimenting, run it on a throwaway VPS rather than one with production data or other apps' secrets sitting in the same environment.
- Review what it's actually being asked to do, especially early on - an agent that can act is worth watching the first dozen times you use it, the same way you'd watch a new hire's first few commits.
None of this is a reason not to run it - it's a reason to run it deliberately.
Pairing it with a self-hosted model
OpenClaw is commonly deployed alongside Ollama so the reasoning behind its responses stays on your own hardware too, rather than calling out to a third-party model provider for every message. Deploy both on the same DeployOS server and they share the private Docker network automatically - no separate networking step. Letta (MemGPT) and Qdrant are worth adding on top if you want it to remember context across sessions rather than starting fresh each conversation.
The result is closer to a personal assistant that happens to run on a $10/month box you own, than a subscription you're hoping doesn't change its terms next quarter.