Example Setups & Model Routing
Real-world setups, tiered models, and keeping costs under control
Real-world setups, tiered models, and keeping costs under control
Once you get past the first install, the next big questions are usually: where should it run, whose accounts does it use, and how do you keep it from costing a fortune? This page is a roundup of what tends to work in practice—reference setups and model-routing ideas from people running always-on agents—so you can steal the bits that fit your situation.
Most people end up choosing one of two paths.
A dedicated machine—a Mac Mini, an old PC, whatever’s always on—keeps OpenClaw separate from your main laptop and your personal login. When something breaks (and it will), you can sit down at that box and debug. The tradeoff: you’re responsible for the hardware.
A VM or cloud instance gives you isolation and portability without a physical box. Nice if you don’t want another device. The catch is that when things go wrong, remote debugging can feel a bit more painful.
A third option is managed cloud hosting (e.g. Clawi.ai): they run OpenClaw for you in the cloud, so there’s no machine or VM to look after—you just connect your messenger and go. Handy if you want zero server ops.
Either way, running the agent on something separate from your daily driver is common—both for security and so it can stay up 24/7 without hogging your machine. More on that in our Security Guide and Deployment docs.
Do you hand OpenClaw your own Gmail and GitHub, or give it its own accounts?
Dedicated accounts mean the agent has its own identity—its own Gmail, GitHub, X, whatever. It’s not you; it works on your behalf. If something goes sideways, your personal identity and data stay out of it. A lot of folks who run agents seriously go this route and still supervise what the agent does (e.g. approving posts before they go live).
Shared credentials are simpler to set up: the agent just uses your accounts. The risk is obvious—any mistake or compromise hits you directly. Fine for experimentation; less ideal if you care about long-term trust and sandboxing.
We go deeper on DM pairing, sandboxing, and trust in the Security section.
If you use one top-tier model for everything—main chat, heartbeats, subagents, memory—bills can climb fast when the agent is always on. The pattern that actually works for most people is tiered routing: put the expensive model where it matters, and use cheaper, faster models for the rest.
Real-world costs are all over the map—from a few tens of dollars to hundreds per month depending on how much you use it. Tune things in Model Providers and Configuration, and use Automation to wire cron and heartbeats so the right tasks hit the right models. If you’d rather not manage tiers manually, plugins that route each request to the cheapest capable model can cut bills further—for example ClawRouter (smart router, 41+ models, pay-per-request with USDC). See Providers for a short note and Skills for discovery.
OpenClaw can run on a schedule—heartbeats every 1, 5, or 30 minutes, or at specific times—so the agent wakes up on its own to check tasks, send briefings, or run routines. That’s when it starts to feel like a persistent team instead of a one-off chat. Set it up via cron and heartbeat; if you’re new to this, start with one daily or hourly job and scale up once you’re happy with the cost.