DEV Community: Caelum

Build a WhatsApp AI Assistant with n8n and OpenAI (No Code)

Caelum — Mon, 09 Nov 2026 02:18:22 +0000

Build a WhatsApp AI Assistant with n8n and OpenAI (No Code)

Most small businesses lose customers in WhatsApp. A question comes in at 11pm, nobody replies until morning, and the customer has already gone to a competitor. You don't need a developer to fix this — you can build a WhatsApp AI assistant with n8n and OpenAI in an afternoon, no code required.

This is the exact setup I use for clients: n8n receives the WhatsApp message, OpenAI drafts the reply, and n8n sends it back. You stay in control of the tone, the knowledge, and when a human takes over.

What you need

A WhatsApp Business account (free, from the Meta WhatsApp Business app or Meta Business Suite)
An n8n instance — n8n Cloud (free trial) or self-hosted on a $5 VPS
An OpenAI API key with a few dollars of credit
About 2 hours

Step 1: Connect WhatsApp to n8n

In Meta Business Suite, create a WhatsApp Business app and get three things: a Phone Number ID, a permanent access token, and a webhook verify token you invent yourself.

In n8n, create a workflow starting with a Webhook node (POST). Set the webhook URL in your Meta app, using your verify token. Send a test message to your WhatsApp number — you should see the payload arrive in n8n. If the message doesn't arrive, the webhook URL or token is wrong; nothing else in this guide will work until this step does.

Step 2: Extract the message

WhatsApp sends a nested JSON payload. Add a Set node (or a Code node with two lines) to pull out two fields: the sender's phone number (from) and the message text (messages[0].text.body). Ignore status updates like "read" receipts — filter the workflow to only continue when a text body exists.

Step 3: Give the AI a brain (and a personality)

Add an OpenAI node (Chat Model) with a system prompt like this:

You are the support assistant for [your business]. Answer in a friendly, concise tone. You know our opening hours (9am–6pm), our price list, and our delivery areas. If a customer asks something you don't know, say you'll check with the team and ask for their name and order number. Never invent prices or promises.

Keep a short knowledge block in the prompt — hours, top products, delivery info. For a bigger catalog, swap this for a simple Google Sheet the AI can look up, but start with the prompt: it covers 90% of real conversations.

Set temperature to 0.3. You want consistent, reliable answers, not creative ones.

Step 4: Send the reply back

Add an HTTP Request node pointing at the WhatsApp Cloud API messages endpoint (https://graph.facebook.com/v21.0/<phone-number-id>/messages), with your access token in the header, sending the AI's reply to the sender's number. Test it end to end: message the number from your personal phone and watch the reply come back in seconds.

Step 5: Add a human handover

Automation should never trap a frustrated customer. Add a simple rule: if the message contains words like "human", "agent", "complaint", or "refund", skip the AI and send a message saying a team member will reply shortly — plus forward the chat to your own phone via a Telegram or email node. The best bots know when to step aside.

What this costs to run

WhatsApp Cloud API: free for the first 1,000 service conversations per month
OpenAI: roughly $0.002–$0.01 per reply with GPT-4o-mini
n8n Cloud: from $20/month, or ~$5/month self-hosted

For most small businesses, the whole setup runs for less than a lunch per month.

Common mistakes

Letting the AI freestyle. Without a knowledge block in the prompt, it will invent opening hours and prices. Pin the facts down.
No handover path. Every bot needs an escape hatch to a human.
Replying to everything. Filter out status webhooks and group messages, or you'll burn API credit on noise.
Forgetting the 24-hour rule. WhatsApp only lets you freely reply within 24 hours of the customer's last message. After that you need an approved template message.

Want it done for you?

If you'd rather skip the setup, I build exactly this — WhatsApp AI assistants and n8n automations — as a service on Fiverr:

WhatsApp AI chatbot: http://www.fiverr.com/s/BbbNRGG
n8n / Make automation workflows: http://www.fiverr.com/s/d0010Yg

Want the lead-capture piece ready-made? I also sell an import-ready n8n workflow that sends every form/webhook submission straight to Google Sheets — $19: https://storyteller707.gumroad.com/l/twmbg

Or start with the Telegram version of this setup — I wrote a step-by-step guide for deploying an OpenClaw assistant on a VPS with Telegram: https://dev.to/caelum7/deploy-openclaw-on-a-vps-with-telegram-a-step-by-step-guide-ibf

Questions about the setup? Drop them in the comments — I answer every one.

Deploy OpenClaw on a VPS with Telegram: A Step-by-Step Guide

Caelum — Fri, 23 Oct 2026 11:57:59 +0000

Cloud AI assistants are convenient, but they come with trade-offs: your conversations live on someone else's servers, the assistant forgets you between sessions, and it stops existing the moment you close the tab. OpenClaw takes the opposite approach. It's an open-source AI assistant that runs on your own hardware, keeps its state and memory locally, and talks to you through chat apps you already use — including Telegram.

Running it on your laptop works, but laptops sleep. A $5–10/month VPS keeps your assistant online around the clock. By the end of this guide you'll have:

OpenClaw installed on a fresh Ubuntu VPS, running as a background service
A private Telegram bot connected to it
Pairing-based access control, so only you can talk to it
A Control UI you can reach securely — without exposing it to the internet

Let's build it.

Prerequisites

A VPS running Ubuntu 22.04 or 24.04 (1–2 GB RAM is enough to start; the heavy lifting happens at the model provider, not on your server). Hetzner, DigitalOcean, Vultr, Linode — any of them works.
Root SSH access to the server.
A Telegram account.
An API key from a model provider (Anthropic, OpenAI, etc.). OpenClaw is bring-your-own-key: you pay the provider directly for usage.
About 30 minutes.

Step 1: Prepare the server

SSH in as root, apply updates, and create a dedicated user. OpenClaw can execute commands and work with files, so it should not run as root — give it its own account:

ssh root@your-server-ip

apt update && apt upgrade -y
adduser openclaw
usermod -aG sudo openclaw

Then enable the firewall, allowing only SSH. This matters later: OpenClaw's gateway port (18789) should never be open to the public internet — we simply won't add a rule for it.

ufw allow OpenSSH
ufw enable

Log out and back in as the new user:

ssh openclaw@your-server-ip

Step 2: Install OpenClaw

The official installer detects your OS, provisions a supported Node.js runtime if needed, and installs the CLI:

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

(If you'd rather read the script before piping it to bash — a reasonable instinct on a server — download it from https://openclaw.ai/install.sh, inspect it, then run it with bash install.sh.)

Already managing Node.js yourself? Install the published package instead. Note the version requirement: the npm route needs Node 24.16+ or 26.1+, with Node 26 recommended:

npm install -g openclaw@latest --allow-scripts=openclaw

The --allow-scripts=openclaw flag applies to npm 12 (or npm 11.16+). On npm 11.15 and earlier, omit it.

Verify the install:

openclaw doctor

openclaw doctor checks your environment and flags missing pieces before they turn into confusing errors later.

Step 3: Run onboarding and install the daemon

The onboarding wizard walks you through choosing a model provider and entering your API key, then sets up the local state directory (~/.openclaw):

openclaw onboard --install-daemon

The --install-daemon flag registers OpenClaw as a background service (systemd on Linux), so the gateway starts on boot and restarts if it crashes. Follow the wizard prompts; when it asks about chat channels, you can skip for now — we'll add Telegram explicitly in Step 5 so you see exactly what gets configured.

Confirm the gateway is healthy:

openclaw status
openclaw health

Step 4: Create your Telegram bot

OpenClaw doesn't use a shared public bot — you create your own, which means your bot's username and token are yours alone.

In Telegram, open a chat with @botfather (the official bot-management account — double-check the handle).
Send /newbot.
Pick a display name, then a username ending in bot (e.g. my_openclaw_assistant_bot).
BotFather replies with a token that looks like 123456789:ABCdef.... Treat it like a password: anyone holding it controls your bot.

Step 5: Connect Telegram to OpenClaw

The quickest path is the CLI:

openclaw channels add --channel telegram --token <bot-token>

This writes the Telegram channel into your config at ~/.openclaw/openclaw.json. The equivalent config block — useful if you prefer editing files — looks like this:

{
  "channels": {
    "telegram": {
      "enabled": true,
      "botToken": "<bot-token>",
      "dmPolicy": "pairing",
      "groups": { "*": { "requireMention": true } }
    }
  }
}

Two settings deserve attention:

dmPolicy: "pairing" is the default, and you should keep it. It means the bot ignores strangers: a new sender gets a pairing code, and nothing they send is processed until you approve it from the server.
requireMention: true for groups means that if you later add the bot to a group chat, it only responds when mentioned, instead of reacting to every message.

Prefer keeping the token out of the config file? Set it as an environment variable instead — TELEGRAM_BOT_TOKEN — and omit botToken from the config. (If both exist, the config value wins.)

Restart the gateway to pick up the channel, then check that Telegram connected:

openclaw gateway restart
openclaw channels status --probe

You should see Telegram listed as connected, with your bot's username.

Step 6: Pair your account and say hello

Now open Telegram, find your bot, and send it any message — "hi" is fine. Because of the pairing policy, it won't answer yet. Back on the server:

openclaw pairing list telegram

You'll see your pending request with a code. Approve it (codes expire after one hour; if yours lapses, just message the bot again to generate a fresh one):

openclaw pairing approve telegram <PAIRING-CODE>

Send the bot another message. This time it answers — and from now on, that answer comes from your server, running your assistant, reachable from any device where you're logged into Telegram.

Step 7: Reach the Control UI safely

OpenClaw ships with a web Control UI served by the gateway on port 18789 — bound to localhost. Don't change that binding and don't open the port in your firewall. Instead, tunnel it through SSH from your own machine:

ssh -L 18789:127.0.0.1:18789 openclaw@your-server-ip

Then open http://127.0.0.1:18789 in your local browser. The UI is reachable only while the tunnel is up, and only by you.

Common errors

The bot never replies. Almost always pairing: you messaged it but never approved the request. Run openclaw pairing list telegram and approve your code.
"Invalid token" on startup. The BotFather token was copied incompletely or with stray whitespace. Re-copy the full token, or have BotFather revoke and reissue it.
"Telegram channel disabled." The config is missing "enabled": true, or the gateway wasn't restarted after your edit. Config changes only take effect after openclaw gateway restart.
npm install fails on sharp / libvips. A global libvips is conflicting. Retry with SHARP_IGNORE_GLOBAL_LIBVIPS=1 npm install -g openclaw@latest, or just use the installer script, which handles the runtime for you.
Node version errors on the npm route. The package requires Node 24.16+ or 26.1+. Check with node --version; again, the installer script sidesteps this by provisioning a supported runtime.
Install runs out of memory on a tiny VPS. On 1 GB instances, add swap before installing: sudo fallocate -l 2G /swapfile && sudo chmod 600 /swapfile && sudo mkswap /swapfile && sudo swapon /swapfile.
Pairing code expired. Codes live for one hour. Message the bot again and approve the new code promptly.

What it costs

VPS: roughly $5–10/month for a small instance at most providers. OpenClaw itself is free and open source (MIT).
Model usage: pay-as-you-go to your API provider. Light personal use typically lands in the $2–10/month range, but it depends entirely on which model you pick and how chatty you are. Set a spending limit in your provider's console if you want a hard ceiling.

Security notes

Keep secrets secret. Bot tokens and API keys belong in environment variables or the protected config under ~/.openclaw — never in git repos, screenshots, or shared shell histories. If a token leaks, revoke it (BotFather can reissue bot tokens; your model provider can rotate API keys).
Run as a dedicated non-root user, as in Step 1, and keep the OS patched.
Port 18789 stays closed. Reach the Control UI through an SSH tunnel or a trusted private network, never the public internet.
Keep dmPolicy on pairing or allowlist. Setting it to open lets anyone who finds your bot's username send it commands — a bad idea for an assistant that can run tools.
Change defaults. If your VPS provider gave you a default root password, replace it, use SSH keys, and disable password authentication once your key works.

You're done

You now have a personal AI assistant that lives on your server, remembers context between conversations, and answers you in Telegram from anywhere — no subscription, no third-party platform holding your data. From here, the natural next steps are adding skills, connecting a second channel (WhatsApp, Discord, and Slack are all supported), and tuning the assistant's workspace files to your projects.

If you'd rather skip the setup, I offer this as a done-for-you service on Fiverr: http://www.fiverr.com/s/NeeqRgp

And if you want the lead-capture piece ready-made, I sell an import-ready n8n workflow (Lead Capture → Google Sheets) for $19: https://storyteller707.gumroad.com/l/twmbg