We set up OpenClaw in steps: agree on the first job, choose the machine, install from the official source, connect one chat channel and a model, connect accounts narrowly, apply security settings, test, and hand off.

We have set up OpenClaw for clients, and the work follows the same shape each time: agree on the first job, choose a machine that will stay on, install OpenClaw from the official installer, connect one chat channel the client already uses, choose a model provider they can pay for and maintain, apply conservative security settings, test the job with real examples, and hand it off with written notes. The install is the short part. Most of the time goes into access, approvals, and making sure the client knows how to supervise the assistant. Product facts here come from openclaw.ai and the official security docs, current as of November 2026.

Step 1: agree on the first job

Before we touch a computer, we write the first job in one sentence and agree on what the assistant may and may not do. "Triage my inbox each morning and draft replies to new inquiries for my approval" is a good first job. "Run my business" is not. The job decides which accounts get connected and which stay out of reach. See What to automate first with AI.

Step 2: choose the machine

OpenClaw runs on macOS, Windows, or Linux, and its state — memory, skills, and configuration — lives on that machine. That means the machine matters. It should:

  • Stay on and awake during the hours you want the assistant working.
  • Be one you control and can physically secure.
  • Not be shared with people who should not reach the assistant.

For some clients that is a dedicated small computer in the office. For others it is an existing machine that is always on. We talk through the trade-offs; we do not insist on new hardware. OpenClaw's official docs describe one trust boundary per gateway, so we also decide early whether this gateway is for one person or for a small team whose members trust each other.

Step 3: install from the official source

We install from the official desktop app or the official command-line installer on openclaw.ai. We do not use unofficial bundles. The desktop apps install the gateway, chat, and setup pieces together, which is usually the simplest path for a business that will not maintain the machine day to day.

We note the version installed and where configuration lives, so the next person who works on it, whether that is us or someone else, does not have to rediscover it.

Step 4: connect one chat channel

OpenClaw officially supports channels including WhatsApp, Telegram, Discord, Slack, Signal, and iMessage. We connect the one the client actually checks all day, and only that one at first. If the team lives in Slack, we use Slack. If the owner runs the business from a phone messaging app, we use that.

We then confirm who is allowed to talk to the assistant. Official defaults help here: unknown direct-message senders get a pairing code rather than having their message processed, and group access is allowlisted. We keep those defaults and add people deliberately.

Step 5: choose a model

OpenClaw is bring-your-own-model. The official site describes hosted, subscription-backed, gateway, and local options. We help the client pick one they can pay for, understand, and keep running. The choice depends on budget, privacy expectations, and how demanding the job is. We write down which provider is used and where its key lives, and we do not paste keys into chat.

Step 6: connect accounts narrowly

Next we connect only the accounts the first job needs: usually one email account and one calendar. Where the tools allow it, we use the narrowest access that still lets the job work. We decide whether the assistant should have full machine access or a sandboxed setup; official docs describe both and treat the choice as yours. For a first business job we lean toward less access.

Step 7: apply security settings and audit

We follow the official security guidance: keep the gateway bound to loopback on a regular host install, keep pairing and allowlists on, and avoid exposing the gateway to the internet without a clear reason. Where it applies, we run the project's own `openclaw security audit` and fix what it flags. Our broader checklist is in Securing an AI agent: basics.

Step 8: write the instructions and approval rule

We write the assistant's instructions with the client: what starts the job, where information lives, what a good result looks like, and when to stop and ask. Then we set the approval rule. Our default is that the assistant may read, sort, and draft, and anything that leaves the business waits for a person. See Approval rules for AI agents.

Step 9: test with real examples

We run the job against real, recent examples from the client's inbox or calendar, with approvals on, and review every result together. We fix instructions where the assistant misunderstood the business. We also test the failure cases: a confusing email, a message from an unknown sender, a request it should refuse.

Step 10: hand off

Handoff includes a short walkthrough and written notes covering how to message the assistant, how to approve or reject a draft, how to pause it, where it runs, which model and accounts it uses, and who to call if something looks wrong. The client should be able to supervise the assistant without us.

What do clients usually ask afterwards?

The most common request after a few weeks is to let the assistant do one more thing on its own, usually sending a type of reply that has been approved many times. We loosen the rule one action at a time. The second most common request is a second job, often follow-up on leads or a daily briefing.

Questions

How long does an OpenClaw setup take?

It depends on the job, the machine, and how many accounts are involved. We scope it on the free discovery call and quote it before starting.

Do I need to buy a new computer?

Not necessarily. The assistant needs a machine you control that stays on when you want it working. Sometimes that is an existing computer. We talk through options.

Can my whole team use one OpenClaw?

OpenClaw supports sharing one gateway with a team, and official docs treat that gateway as a single trust boundary. It suits a team whose members trust each other. If people need isolation from each other, official guidance is to split gateways.

Do you only set up OpenClaw?

No. We have also set up Hermes Agent and Grok Bot for clients, and we recommend what fits.