Before you start: prerequisites
OpenClaw runs on your own hardware, so you need a machine and a few accounts.
- A machine to run it on: your laptop, a Mac mini, a VPS or a server.
- A model provider API key (your choice of LLM provider), or a local model.
- At least one channel to talk to the agent, such as a chat app.
- The runtime and tooling the project requires, per the official OpenClaw docs.
Install OpenClaw
The high-level flow is the same on any machine. Follow the official instructions for the exact commands.
- Get OpenClaw from its official source, the project repository.
- Install its dependencies following the official instructions.
- Create your configuration: where state lives, which channels and skills are enabled.
- Pin your version so an update does not change behavior unexpectedly.
Configure your model and channels
OpenClaw is model-agnostic, so you choose your provider.
- Add your model provider key, or point it at a local model.
- Connect a channel so you can send goals and get results back.
- Enable the skills and integrations the agent needs for its job.
- Store credentials carefully: they give the agent real access.
Run your first agent
Start OpenClaw, send it a goal from your channel, and watch it work: it reasons, calls the skills you enabled, and reports back. For a single personal agent on your own machine, you are done.
From a local install to production
Installing one OpenClaw is the easy part. The hard part starts when it has to run unattended, recover when it crashes, survive version updates, and scale to many agents without you babysitting each one. At that point you are operating infrastructure, not running an app.
A managed runtime removes that work: provisioning in seconds, automatic recovery, a versioned filesystem, integrations ready on day one and protection when agents spike together. Molted runs OpenClaw this way, so you skip the production setup entirely.
One agent
Easy to babysit.
A fleet, by hand