How to set up OpenClaw and 15 jobs it can take off your team

The first useful thing I asked OpenClaw to do was boring. Go through last week's supplier emails, work out which promised documents never arrived, and write me the chase list.

Forty seconds. The list was right. Somebody had been doing that by hand every Monday for a year.

That is the appeal. OpenClaw is not a chat window that gives advice. It runs on a machine you control, it reads files, sends messages, browses the web and runs tasks on a schedule, and you talk to it through WhatsApp, Telegram or Slack like a colleague. It remembers what you told it last week.

This is the setup guide plus 15 jobs it does well, three for each of five departments. Read the setup, then skip to your own department. If the word terminal makes you tense, the only part that needs one is three lines and I have written them out.

Three decisions before you start

Which machine. A laptop or small server you would be comfortable wiping. Not the one holding the payroll file.

Which account. Ideally not your personal login. Create a separate user with access to only what this assistant needs.

Which model. You need an API key from a provider such as Anthropic, OpenAI or Google. That is the part that costs money, and for these jobs it is a small number, not an enterprise contract.

Installing it

You need Node.js 22.22.3 or newer, with 24 recommended. The installer does the rest.

On macOS or Linux:

curl -fsSL https://openclaw.ai/install.sh | bash

On Windows, in PowerShell:

iwr -useb https://openclaw.ai/install.ps1 | iex

Then run the wizard, which asks for your model provider and sets it running in the background:

openclaw onboard --install-daemon

Check it is alive and open the dashboard:

openclaw gateway status
openclaw dashboard

Get a reply in the chat box and the technical part is over. It usually takes five minutes.

Connecting it to where you already talk

The point is that you message it from your phone, not from a dashboard.

Telegram is fastest because it needs only a bot token. Slack, Discord, WhatsApp, Signal and Teams work too, with a few more steps.

One setting matters more than the rest. In the config, dmPolicy decides who may talk to your assistant. Set it to pairing, which requires a one-time code, or allowlist with named accounts. Never leave it on open. An assistant holding your files and your email that answers to strangers is not an assistant.

Teaching it who you are

Most people skip this, and it is the difference between a novelty and something you use daily.

OpenClaw keeps a workspace folder with a few plain text files. AGENTS.md is the important one: it is where you write, in ordinary English, how it should behave. MEMORY.md holds the durable facts, and it writes daily notes as it goes.

Write what a new assistant would need on day one. How you like replies. Who the important people are. What your project names mean. Which folders it may touch. What it must never do without asking.

Be specific. "Keep replies under five lines unless I ask for detail" changes every future conversation. "The supplier register in Finance/Suppliers is the source of truth over anything in email" prevents a whole class of confident nonsense.

Then write your red lines in the same file. Never send anything to a customer without showing me. Never delete anything. Ask before touching a config file.

Twenty minutes on that file pays back inside a week.

The five minutes that keep it out of trouble

OpenClaw went viral in January 2026 and by mid-February researchers found over 135,000 instances sitting exposed on the public internet, because most people never change a default. The bugs were patched quickly. The defaults are your job.

Turn sandboxing on rather than giving it the run of the machine. Use the deny list to block anything it does not need yet, such as the browser or shell commands. Keep logging on with sensitive values redacted. And keep the skills list short.

That last one matters most. Community skills are code written by strangers running with your permissions, and a security firm found several hundred malicious ones on the public hub this year. Adding a skill is installing software on a work machine, so treat it that way.

Engineering

The overnight digest. At 07:30 it reads the alerts, error rates and deploys since yesterday evening and posts one message: what broke, what recovered on its own, what still needs a person. Read-only access to monitoring, write access to one channel. It is the safest possible first job and it converts sceptics, because everyone can see whether the summary was right.

Vulnerability triage that is actually readable. Once a week it checks your lockfiles against published advisories and opens a single issue listing only what genuinely reaches your code, with the path that gets there. The problem was never a lack of alerts. It was that 200 of them arrive at once and nobody can tell which 3 matter.

The incident timeline. After something breaks, it assembles the draft postmortem from the Slack thread, the deploy log and the alert history, in order, with timestamps. The engineer who was up until 02:00 edits it instead of reconstructing it. Most postmortems are late because nobody wants to rebuild the story from scratch.

The CEO

Where you are the bottleneck. Monday at 07:30, a list of decisions waiting on you specifically: threads where you are the last unanswered person, approvals sitting in a queue, things somebody flagged twice. Not a summary of your inbox. A list of what stops if you do nothing this week.

The board pack pre-read. It pulls the same numbers from the same places every month, drops them into your standard template with the change against last month, and notes which ones moved more than usual. You write the narrative, which is the part that needs you. You stop spending a Sunday assembling figures.

The quiet competitor watch. Daily, it checks a named list of competitors for changes to pricing, positioning and job postings, and messages you only when something actually changed. Job postings are the underrated one. A company hiring three compliance people is telling you what it is about to do.

Customer service

Drafts, never sends. It reads each incoming ticket, finds the relevant order or account, and writes a suggested reply as an internal note. A human presses send. First response time falls because the reading and the lookup are done. Risk stays near zero because the sender is still a person. If you ever let it send directly, EU rules in force since 2 August 2026 require telling the customer they are dealing with a machine.

The repeat question report. Every Friday it groups the week's tickets and tells you the five things customers asked most, in the words they actually used. That list is your help centre backlog, your product bug list and your onboarding gaps, and it is usually the highest value document nobody in the company produces.

The broken promise tracker. It scans conversations for commitments your team made, such as a refund by Friday or an engineer looking on Monday, and flags any that passed their date without happening. Almost every serious complaint starts as a small promise nobody tracked.

Sales

The 30 minute pre-meeting brief. Half an hour before each meeting in your calendar, one screen arrives: who you are meeting, the last three interactions, anything currently open with support, and what you promised last time. Walking into a call knowing about their open ticket is the difference between a good meeting and an ambush.

CRM notes without the CRM chore. After a call you send it three lines or a transcript. It writes the proper account update, sets the next step, and chases you if that next step has no date on it. Sales people do not avoid the CRM because they are lazy. They avoid it because it is the fourth time they have written the same thing.

The gone quiet list. Weekly, the deals where nothing has happened in 14 days, each with a draft re-engagement message that references the last real conversation rather than pretending to check in. Most pipeline does not die from rejection. It dies from silence nobody noticed.

Procurement

The renewal radar. It reads the contracts folder, extracts renewal dates and notice periods, and warns you 90 days ahead with the deadline and the owner. Auto-renewal is the most expensive thing in most software budgets, and it happens because a date sat in a PDF nobody opened.

Chasing documents. Certificates, insurance, compliance paperwork. It compares what should be on file against what is, drafts the chase messages, and you send them. This was my first job for it and it is still the one I recommend starting with: dull, relentless, and a wrong answer is obvious in seconds.

Invoice against contract. Each month it checks the invoices against the agreed rates and flags the lines that do not match, along with anything billed that nobody recognises. It pays for itself the first time a vendor's per-seat price quietly changes at renewal.

One to avoid, since people ask: do not point it at CV screening or anything deciding who gets a job, a loan or a place on a course. That sits in the high-risk category of the AI Act, and it is not a first project.

Keeping it honest

For the first fortnight, read the logs. Not because it will go wrong, but because you will learn what it actually does, which is never quite what you assumed.

Then ask one question of every job you give it: would I have noticed if it had been wrong? If the answer is no, that job belongs with a human until you have a way to check.

And keep it updated. This project ships fast, and an old version is the one with the known holes.

Where to start

Pick one. Not fifteen. The temptation with a list like this is to try five at once and end up trusting none of them.

Take the one on your department's list that is dull, repetitive and instantly checkable, and run it for two weeks before adding anything. For most teams that is the chase list or the overnight digest, which is why both are on there.

I write up what I am learning from running these every fortnight in Operating Notes.