Quick Answer
Your customer messages at 10pm with a broken boiler. By the time you check the phone in the morning, they have already rung someone else. OpenClaw fixes that. It is a free, open-source AI agent that runs on a cheap VPS, plugs into the WhatsApp Cloud API, and talks to your Google Calendar. Customer asks for a Tuesday slot, the agent checks your diary, books the appointment, sends the confirmation, and adds the event. At TrainAR we have run this stack with a handful of trades businesses and the maths is simple. Trades lose 30 to 50 percent of after-hours leads. This setup captures the lot, for under thirty quid a month. Now compliant with the January 2026 Meta rules on AI agents, because the agent is task-specific, not open-ended.
Table of Contents
- What this stack does
- Why 24/7 cover matters for trades
- How OpenClaw, WhatsApp and Google Calendar fit together
- What you need before starting
- The setup, step by step
- Guardrails: keeping the agent honest
- Staying inside Meta's 2026 AI rules
- Cost: DIY versus the receptionist services
- My verdict
- What the automation community is saying
- Recommended videos
- Frequently asked questions
What this stack does
OpenClaw
WhatsApp Business
Google Calendar
Claude (the LLM)
Picture the most common after-hours scenario. A homeowner gets in from work, kettle on, sits down, and realises the boiler is making a noise it should not. They open WhatsApp because that is where they message everyone, type out a description, hit send. Then they wait. If you reply in the morning, half the time they have already booked somebody else who answered at half past ten the night before.
This stack puts a polite, accurate, scoped AI agent in between that customer and your sleep. OpenClaw is the open-source AI agent platform that crossed 250,000 GitHub stars in February 2026, runs on your own server, and is free under the MIT licence. WhatsApp Business is the channel customers actually use. Google Calendar is your diary. Claude (or whichever LLM you point OpenClaw at) is the brain. You wire them together once, pay for the LLM tokens by usage, and let the agent qualify the lead, check the diary, propose a slot and book it.
The agent does four jobs, in this order. It greets the customer with the legally required AI disclosure, asks the questions you would ask (what is the issue, what is the postcode, is it urgent), checks your Google Calendar for the next available slot that fits, and books the appointment with a calendar event and a WhatsApp confirmation. Anything outside those four jobs, like trying to diagnose the boiler over WhatsApp, the agent refuses and offers to call you. That refusal is not a weakness, it is the thing that keeps you inside Meta's January 2026 chatbot rules.
Why 24/7 cover matters for trades
I have lost count of the conversations where a trades business owner tells me they get most of their enquiries on the weekend or in the evening, then in the same breath complains that they cannot find good admin staff. The maths usually goes unsaid. If you take a typical heating firm doing twelve enquiries a week and roughly half come outside hours, you are dropping six potential jobs every Monday morning. Even a 30 percent conversion rate on those means two jobs a week lost. That is at least £1,500 of revenue gone before you have even opened the laptop.
The benchmark from Checkatrade's own research is brutal here: the first tradesperson to respond often wins the job, and you should be replying within an hour during working hours. Outside hours that drops to within minutes if you want to win. Industry studies put the average missed-call rate for home service businesses at around 27 percent. WhatsApp data tells the same story from the customer side: 73.3 percent of consumers across 22 markets now prefer messaging a business when they make contact, and in the UK specifically 73 percent of WhatsApp users message brands on it every day.
The cost of an after-hours gap
A two-engineer plumbing firm receiving 120 written enquiries a month, with 60 percent landing outside hours and a realistic 35 percent conversion on captured leads, leaves around 25 jobs a month sitting in voicemail. At an average ticket of £220, that is £5,500 of monthly revenue going to whoever answered first. The agent costs you under thirty quid.
The other thing nobody talks about is the emotional cost of always being on. Most trades business owners I work with check WhatsApp last thing at night and first thing in the morning because they are scared of missing a job. That is not a sustainable way to run a business, and it is not a sustainable way to be a parent or a partner either. A well-scoped agent gives you back the evenings. It is not a replacement for you, it is the part of the job that should not need you.
How OpenClaw, WhatsApp and Google Calendar fit together
The shape of the system is genuinely simple, and that is by design. WhatsApp Cloud API takes the inbound message and forwards it as a webhook to your OpenClaw instance. OpenClaw runs the message through the LLM with a tight system prompt that defines the agent's job. The agent has access to two tools: a Google Calendar tool for reading and writing your diary, and a WhatsApp reply tool for sending messages back. Persistent memory in OpenClaw means the agent remembers what the customer said three messages ago, so it does not ask for the postcode twice.
The four moving parts to wire up are: the WhatsApp Cloud API webhook (free up to a point, with conversation pricing after that), the OpenClaw container running on a small VPS, the Google Calendar OAuth credentials, and the LLM API key (Claude Sonnet 4.6 is the sensible default for cost and reliability on JSON responses). If you are already running n8n for other automations, you can sit n8n in front of OpenClaw and use it as the webhook router. We built our first version with n8n in the middle and it added flexibility, but you do not need it.
Why we use Claude not GPT for this
OpenClaw is model-agnostic, but Claude Sonnet 4.6 wins on three counts that matter here: it follows JSON output instructions more reliably than GPT-5 in our testing, it refuses out-of-scope requests cleanly rather than waffling, and at the token volume of a typical trades business the monthly bill comes out 30 to 40 percent lower. If you have an OpenAI credit sitting unused, GPT-5 also works and the prompt does not change.
The whole thing runs on a VPS that costs about a fiver a month. Hetzner CAX11 is the one I keep recommending, it is an ARM box with 2GB RAM and it runs OpenClaw and n8n together with room to spare. The LLM token cost is the variable one. A normal booking conversation uses about 800 to 1,500 tokens. At current Claude Sonnet pricing that is fractions of a penny per conversation, so even at 500 conversations a month you are looking at under £15 in LLM costs.
What you need before starting
This is not a no-code setup. You do not need to be a developer, but you do need to be comfortable in a terminal for about 20 minutes and able to follow a YAML file. If you have ever installed something on a VPS using a one-liner, you can do this. If your eyes glaze over at the word "terminal", get whoever does your IT to follow this with you in an hour.
| Requirement | Details | Cost |
|---|---|---|
| VPS (cloud server) | Hetzner CAX11, Hetzner CX22, or any 2GB ARM/x86 VPS. Ubuntu 22.04 or 24.04. Used to host OpenClaw 24/7. | £3-5/mo |
| WhatsApp Business Cloud API access | Set up through Meta Business Suite. You need a Facebook Business Manager account and a phone number you can verify by SMS or call. | Free entry, pay-per-conversation after |
| Google Workspace or Gmail account | The Google Calendar the agent will read and write to. Personal Gmail works but Workspace is tidier for business use. | Free-£6/user/mo |
| LLM API key | Anthropic API key for Claude Sonnet 4.6, or OpenAI key for GPT-5. Used by OpenClaw to drive the agent's responses. | ~£5-20/mo at typical volume |
| A domain name with HTTPS | WhatsApp Cloud API webhooks require HTTPS. A subdomain like agent.yourtradesbusiness.co.uk works fine. Use Caddy or Cloudflare Tunnels for the certificate. | £10/yr (or free with Cloudflare) |
You will also want to write down, before you start, the seven or eight types of enquiry your business actually gets. Boiler not heating water. Boiler making noise. New install quote. Service due. Landlord gas safety check. Bathroom leak. Outside tap. Whatever yours are. The agent only needs to handle the ones that are routine bookings. Anything outside that list, it passes to you. Defining the in-scope list before you build is the part that takes 30 minutes of thinking and saves you days of trying to make the agent handle everything.
The setup, step by step
I am going to walk you through this at the level of detail that actually helps. There are too many tutorials that skip the bits where you get stuck. The total time is about 90 minutes if you are doing it for the first time, 30 minutes if you have done a VPS install before.
Spin up the VPS and install OpenClaw
Create a Hetzner CAX11 (or any 2GB ARM/x86 box), pick Ubuntu 24.04, and SSH in. Run the official OpenClaw one-liner: curl -fsSL https://openclaw.ai/install.sh | bash. The installer handles Node.js 22 and walks you through onboarding. Pick Anthropic as your AI provider and WhatsApp as your chat platform when prompted.
Set up your WhatsApp Cloud API access
Go to Meta Business Suite, create a Business app, add the WhatsApp product, and follow the verification steps. You will need to add a phone number (use a dedicated one, not your personal mobile) and verify it. Grab the permanent access token, phone number ID and WhatsApp Business Account ID. These three values go into OpenClaw's WhatsApp channel config.
Point the WhatsApp webhook at OpenClaw
In Meta's app dashboard, set the webhook URL to https://agent.yourdomain.co.uk/whatsapp/webhook (the URL OpenClaw exposes after install). Subscribe to the messages event. Verify the webhook using the token you set in OpenClaw's config. If verification fails, it is almost always because the domain is not properly HTTPS. Use Caddy or a Cloudflare Tunnel to terminate TLS.
Connect Google Calendar through OAuth
Create a Google Cloud project, enable the Calendar API, and create an OAuth client (Desktop application type). Download the credentials.json file. In OpenClaw, install the Google Calendar skill from ClawHub and point it at the credentials file. The first run will open an OAuth consent page, sign in with the calendar account, copy the verification code back to the terminal. Test by asking OpenClaw to list this week's events. If it does, the calendar tool is live.
Write the agent's system prompt
This is the part that decides whether your agent feels professional or feels like a chatbot. The prompt has to do three things: state the agent's job (book appointments for a specific trades business), define the in-scope enquiries (the list you wrote down earlier), and give a clean refusal pattern for anything else. We have a template you can copy below. Drop it into ~/.openclaw/skills/booking-agent/instructions.md.
Test with a friend before you go live
Switch your WhatsApp Business account to test mode, give a friend your test number, and have them book a slot. Watch the conversation, watch the calendar event get created, watch the confirmation come back. Then ask the friend to try to break it. Ask if it can quote for a new bathroom. Ask if the engineer is qualified for commercial gas work. Anything where the agent should say "I will get the team to call you", make sure it does. Iterate the prompt until the refusals feel like a polite receptionist, not a robot.
A starting system prompt for the agent
"You are the booking assistant for [Business Name], a [trade] business in [town]. You only book routine appointments for the following enquiries: [list]. For anything else, including pricing for non-routine work, technical diagnostics, complaints or commercial enquiries, politely apologise and say a member of the team will be in touch within one working day. Always disclose at the start of a new conversation that you are an AI assistant. Confirm the customer's postcode, the nature of the issue, and a preferred date or time. Check the calendar before proposing a slot, and only propose slots that are actually free. Once a slot is agreed, create the calendar event and send a written confirmation that includes the date, time, address and contact number for the business."
Guardrails: keeping the agent honest
The reason most trades businesses give up on chatbots is that they hallucinate. The agent invents prices, agrees to things it should not, or sends a customer a slot that does not exist. OpenClaw has three guardrails that solve this when you configure them properly, and you should configure all three.
The first is tool-only actions. The agent cannot create a calendar event without calling the Google Calendar tool, which physically reads the diary first. That removes the "I have booked you in for Tuesday" hallucination, because if Tuesday is not free, the tool refuses. The second is the structured response format. The system prompt asks the agent to return its replies as JSON with intent labels (book, ask, refuse, escalate). If a reply does not parse, OpenClaw kicks it back to the LLM with the format reminder. The third is the human-handover trigger. Configure OpenClaw to escalate to you (via SMS or email) any conversation that includes keywords like "urgent", "emergency", "leak", or any conversation longer than ten messages.
Never give the agent your prices
Do not put your price list in the system prompt. The temptation is huge because customers always ask, but prices change, jobs vary, and an AI agent committing you to a number is a liability. The clean answer is "I cannot quote on the spot, but I can book a free 15-minute visit so we can give you a fixed price." That answer wins more jobs than a generic ballpark anyway.
The fourth thing, which is not strictly a guardrail but is good operational hygiene, is daily review. OpenClaw can dump the day's conversation transcripts to a Google Sheet or a Slack channel. Spend five minutes reading them each morning. You will find one or two each week where the agent did not quite get it right, and you adjust the prompt. After two or three weeks of this the agent is doing its job with no edits and you stop checking.
Staying inside Meta's 2026 AI rules
On 15 January 2026, Meta changed the rules. Open-ended AI chatbots on WhatsApp are now banned. ChatGPT and Perplexity were the casualties everyone heard about. What did not get the same coverage is that task-specific AI agents, for shopping, customer support and appointment booking, are explicitly allowed. The whole reason this setup is worth doing in 2026 is that it is exactly the kind of agent Meta wants on the platform.
To stay compliant you need three things in place. First, your agent's opening message must include an AI disclosure ("Hi, I am the booking assistant for [Business]. I am an AI agent and can help arrange your appointment."). Second, the agent must have a defined purpose. The booking-only scope is the defined purpose, and the refusal pattern is what proves it. Third, you must provide human oversight, which the daily transcript review covers. There is no certification process, but if Meta flag your account because somebody complained, having the disclosure and the scoped prompt is what gets the account reinstated.
The WhatsApp Cloud API pricing changed in 2026 too. The old "first 1,000 conversations free" tier is gone. Now you pay per delivered message in four categories: Marketing, Utility, Authentication and Service. For a booking agent, almost every conversation is Service or Utility, both of which are cheaper than Marketing and most are free if they happen inside the 24-hour customer service window. For a typical trades business that means the WhatsApp bill itself is rarely above a few pounds a month. The LLM tokens cost more than the WhatsApp transport.
Cost: DIY versus the receptionist services
The honest comparison here is not OpenClaw versus a free WhatsApp setup. It is OpenClaw versus the alternatives, which are an out-of-hours phone answering service, a managed AI receptionist platform, or doing nothing. Let us put them side by side.
| Option | Typical monthly cost | What you get | Best for |
|---|---|---|---|
| OpenClaw self-hosted (this guide) | £25-35 | 24/7 WhatsApp coverage, calendar booking, customised to your enquiries, full control over the script | Trades businesses with one technical person on the team, or a TrainAR-style support partner |
| Managed AI receptionist platform | £199-499 | Voice and WhatsApp coverage, set up for you, less flexibility, often US-priced in dollars | Businesses with budget but no time |
| Human answering service | £200-600 | Real person, no booking integration, message taken and passed on | Businesses where customers genuinely want a human voice |
| Doing nothing | £0 direct | You keep losing 30-50 percent of after-hours leads, you stay on call mentally every evening | Nobody, really |
The DIY route is the cheapest by some distance, but it costs you the setup time, around 90 minutes for the first build, plus 30 minutes a week of transcript review for the first month. The managed platforms are fine if you would rather pay for it, but in our experience they hit a ceiling when you want to change something specific to your business. The OpenClaw route lets you adjust the prompt the same evening you spot the problem in a transcript. For a trades business that grows or changes its services often, that flexibility is worth the extra hour of setup.
If you already have an automation stack running, this is genuinely additive. Pair it with the workflow in our n8n referral engine guide and you have a system that captures, books and rewards customers without you opening the laptop. Or sit it alongside the WhatsApp routine in our piece on heating engineers and the quote-stage drop-off for follow-up after the booking. If you are still thinking about whether AI belongs in your business at all, the AI-generated marketing content guide is the gentler starting point.
My verdict
Build it, run it for a fortnight, then decide
The honest test for any tool in a trades business is whether you would miss it after a fortnight. Run OpenClaw on WhatsApp for two weeks, read the transcripts, count the jobs it booked while you were asleep. If it has not paid for itself by day ten, switch it off and you are out a Hetzner box and an evening. In every case I have seen, it pays for itself in the first week. A trade is a craft, and customer response is part of the craft. This is the bit that should not need you at midnight.
What the automation community is saying
Recommended videos
OpenClaw Full Setup Tutorial: AI Agent on WhatsApp in 15 Minutes (2026)
Thetips4you
OpenClaw AI Agent Tutorial: Use OpenClaw with WhatsApp & Run Skills Locally
Code With Aarohi
OpenClaw AI Agent: Build Your Own WhatsApp AI Assistant
Unity Coding
Automate Appointment Booking on WhatsApp: n8n x WhatsApp
Krish | AI Automation
This WhatsApp AI Agent Automatically Books Appointments For You
Matt Penny | Applied AI
How I Built my WhatsApp Appointment Booking Workflow with N8N
Zero Point Engineer
Frequently asked questions
No, and you would not want to. The agent needs a WhatsApp Business Cloud API number, which is separate from personal WhatsApp. Use a dedicated business number you can verify. Some businesses use a Google Voice or Vonage number specifically for this; others get a second SIM. Either is fine.
You read the transcripts each morning for the first month. When you spot something the agent should have handled differently, you adjust the system prompt that evening. After two or three weeks of small tweaks, the error rate drops to near zero. The escalation rule also catches the cases where the agent itself flags something needs a human.
Yes, if you configure it as a task-specific booking agent with the opening AI disclosure and the refusal pattern. The open-ended chatbots that got banned were general-purpose assistants like ChatGPT on WhatsApp. Appointment booking is one of the use cases Meta explicitly allows.
About 90 minutes the first time, broken down as: 15 minutes on the VPS install, 30 minutes setting up the WhatsApp Cloud API access in Meta Business Suite (the slowest part), 20 minutes on Google Calendar OAuth, and 25 minutes writing and testing the system prompt. If you have done a VPS install before, halve all of that.
Yes, and they should. The first message the agent sends in any new conversation discloses that it is an AI assistant for the business. In our experience customers do not mind, they just want their booking sorted. The disclosure is also required by Meta's rules and by good practice generally.
Put n8n in the middle as the webhook router. It gives you a visual interface for tweaking the routing logic and lets you fan out the same WhatsApp message to other systems, for example logging the conversation to a Google Sheet or pushing the booking into your job management software. Our other guides cover the n8n side in detail.
It should not try to. The escalation rule should trigger on words like "urgent", "emergency", "leak" or "no heating" and forward the conversation to you immediately via SMS or push notification. The agent's job is to handle routine bookings, not to triage a real emergency. That is a human decision.
For a typical trades business doing 100 to 300 WhatsApp conversations a month: VPS £3-5, Claude API £5-15, WhatsApp Cloud API typically £0-5 (most conversations fall inside the free service window), and your Google Calendar is already free. Total around £15-25/month. The biggest variable is LLM token use; if you send the agent a lot of context, it climbs.
