DEV Community: Darya Nazarava

How I built a WhatsApp AI Agent by prompting Claude (official API, no canvas)

Darya Nazarava — Mon, 15 Jun 2026 14:07:57 +0000

If you have ever tried to add a WhatsApp bot to a product, you know the work is rarely the bot. It is the WhatsApp Business setup, the webhook plumbing, and the per-platform glue you have to keep alive every time Meta changes something.

I wanted to skip all of that, so I built a working WhatsApp AI chatbot by describing it to Claude in plain language and letting an agent ship it for me. It runs on the official WhatsApp Business API, it answers real messages, and I did not draw a single node by hand.

This is a walkthrough of how that works: what the bot does, how Claude builds it over MCP, what is actually happening underneath, and what it costs to run.

Why build on the official WhatsApp API at all?

A lot of "instant, no approval" WhatsApp automation connects through the WhatsApp Web protocol. It is fast to start, but it is unofficial, it goes against Meta's terms, and numbers get banned for it, sometimes in waves. That is fine for a throwaway internal script. It is risky for anything customer facing or anything you bill for.

The official WhatsApp Business API does not get your number nuked, and it is the only route to calling, Flows, and verified templates.

The usual catch is the setup: the WhatsApp Business Account, number registration, and ongoing maintenance. The approach below takes that part off your plate, so you keep the official stack without doing the bureaucratic legwork yourself.

What the Agent actually does

Before the build, here is the target. The example bot is a customer support agent that handles an inbound WhatsApp message end to end:

A customer messages the WhatsApp number, and the bot greets them. If the sender is already a saved contact, it greets them by name.
The bot uses an AI node (running on Claude) to read the message and classify intent: a buying question, a general question, or something that needs a human.
A general question goes back to the AI node, which answers from your own context and then waits for the next reply to keep the conversation going.
A buying question gets routed to your pricing page, tags that contact as a hot lead, and fires a webhook with the full conversation to your CRM or database.
Anything the bot cannot handle is handed off to a human operator.

How to build the workflow using Claude

You only need two tools:

A paid Claude subscription (or any AI agent of your choice)
A Zernio account (to get your WhatsApp agent up and running in less than an hour)

Here's the full workflow:

1. Connect a WhatsApp number

You'll need a phone number for your bot to answer on. You can connect your own WhatsApp Business number or buy a dedicated one through a third party platform.

Before building, make sure the number is connected, verified, and approved in the settings.

If you don't have a number yet, Zernio can handle the purchase and verification for you, letting you set one up in 50+ countries directly from the platform.

2. Add the MCP server as a connector

MCP (Model Context Protocol) is the standard way to give an AI agent a set of tools it can actually use. Zernio's MCP server is a hosted endpoint that exposes the WhatsApp toolset (creating workflows, sending messages, managing numbers, and more) in that format.

Connecting it is what lets Claude do the building instead of you: the agent reads the available tools, then calls them to assemble and publish the workflow. It is hosted, so there is nothing to run or self-host.

For Claude Code, install this plugin:

/plugin marketplace add zernio-dev/zernio-claude-plugin /plugin install zernio@zernio

For Claude desktop, add the connector https://mcp.zernio.com/mcp .

Once connected, the agent can see the WhatsApp tools it needs to create and publish a workflow.

3. Describe the bot in plain language

This is the whole build step. Tell the agent what you want by outcome, not by node:

Build a WhatsApp bot for my number. When someone messages, greet them by name if they are a saved contact. Use Claude to decide if the message is a buying question or a general question. Answer general questions from our docs. For buying questions, reply with our pricing page, tag the contact as a hot lead, and send the full conversation to our webhook. Hand off to a human if it cannot help.

One thing I learned quickly: describe the goal, not the graph. If you micromanage with "add a condition node here," you fight the agent. If you describe the behavior you want, it maps intent to the right nodes more reliably than you would expect.

4. The agent assembles and publishes it

The agent translates that description into a node graph and publishes it to your number.

You can always ask Claude to make changes later. Every update is saved as a new version inside Zernio, so you can easily restore a previous version if something breaks.

5. Text the number and watch it run

Send a message to the number. A greeting comes back. Ask a product question, and the AI node answers in a few seconds, then waits for your next reply. Say you want to buy, and the bot sends the pricing link, tags your contact as a hot lead, and posts the conversation to the webhook.

Every run is logged with timestamps for each node, so you can open any conversation and see exactly which path it took.

What the agent shipped: the nodes underneath

You do not have to know the node set to build this, but it helps to understand what the agent is wiring together. These are the building blocks a WhatsApp workflow is made of.

Node	What it does
Trigger	Starts the flow on an inbound WhatsApp message
Send message	Sends text, media, or a template
Wait for reply	Pauses until the customer responds
Condition	Branches the flow with if/else logic
AI	Sends the conversation to your LLM (OpenAI, Anthropic, Gemini, Mistral, or Groq) and returns the reply, using your own key
Webhook	Calls your database, CRM, or any external service
Set tag	Tags the contact for later filtering or follow-up
Handoff	Transfers the conversation to a human operator

The AI node is the part that turns a flowchart into something that feels like a conversation. It runs on your own LLM key, so you choose the model and pay the provider directly.

If the agent gets a branch slightly wrong, you are not stuck. The same workflow opens in a visual canvas where you can drag, rewire, and republish, and version history keeps every revision so you can restore an earlier one. The agent gives you a fast first draft. The canvas gives you fine control. The API gives you a way to ship the same thing inside your own product. Same workflow, three ways to reach it.

What does a WhatsApp AI chatbot like this cost to run?

This is the part most tutorials skip, and it is the part that decides whether the project is worth it. The build itself is close to free. The running cost is mostly WhatsApp's own fees.

The builder - Zernio. Included, no per-bot or per-contact charge.
Phone number. Around $2/mo for most countries, or bring your own.
WhatsApp messaging (Meta). Since Meta's 2025 change, customer-initiated service replies inside the 24-hour window are free, worldwide. A support bot answering inbound messages pays Meta nothing for those replies. You only pay when you start a conversation with a WhatsApp template, where marketing runs roughly $0.01 to $0.14 per message and utility template is lower, both varying by country.
LLM tokens. Billed by your model provider directly.

So an inbound support bot is, in practice, the number plus pennies in tokens. The cost only climbs if you push outbound marketing templates at volume. That is worth knowing before you pick a tool (like Twilio) that bills you on top for something Meta already charges for.

Closing

From connecting a WhatsApp number to having a bot that classifies messages, routes hot leads, and sends data to a webhook took less than an hour.

Zernio handles the official WhatsApp API setup and maintenance, while the MCP server lets Claude build the workflow for you. That means you don't have to deal with business verification, webhook setup, or other technical work that usually takes days.

It also changes the pricing story. Many WhatsApp chatbot tools charge per contact and add their own monthly fee on top of Meta's messaging costs, so a "$29/month" tool can quickly become much more expensive. With Zernio, the workflow builder is included, and you mainly pay Meta's official messaging fees.

The easiest way to see it is to try it yourself. Connect a number, describe the bot you want, and start with a simple FAQ flow. Then watch it run and ask Claude to improve it step by step.

If you end up building something cool with the WhatsApp API, I'd love to hear what you created.

Auto-Post to 9 Platforms with Claude Code and Zernio

Darya Nazarava — Wed, 20 May 2026 10:13:40 +0000

What you'll build: An agent that analyzes video or images and generates platform-optimized captions, hashtags, and titles—then publishes to 9 social platforms at once.

The problem

Manual posting kills momentum. You format content for Instagram, then rewrite everything for TikTok. Again for LinkedIn. Again for X. A single video turns into a 2-hour job.

With this setup, one prompt becomes nine scheduled posts - captions, hashtags, and format conversions all handled automatically.

What you need

A Zernio account (free tier: 2 platforms)
Claude Code installed
30 minutes to set up; 5 minutes per post after

Step 1: Connect platforms in Zernio

Zernio is the layer that handles the actual platform integrations. Without it, you'd need to build separate OAuth flows, API integrations, media handling, and rate-limit management for each platform (months of work before Claude could publish a post).

With Zernio, that's already done. Claude talks to one API; Zernio handles the cross-platform part.

Log in at zernio.com
Go to your profile and click Connect platform for each account you use (Instagram, TikTok, X, YouTube, LinkedIn, Threads, Pinterest, etc.)
Authorize Zernio for each
Verify each shows in your dashboard

Step 2: Get the Zernio Publishing skill

The skill includes platform-specific guidelines for 9 platforms -hashtag rules, character limits, format specs, tone cues. You're not training Claude from scratch.

Download the skill from the Zernio library GitHub repo
Extract the zip
You'll get:

zernio-library-skills/
├── .claude/
│   └── skills/
│       ├── zernio-publish/                    ← the publishing skill
│       │   ├── SKILL.md
│       │   ├── reference/                       deep API + per-platform docs
│       │   └── templates/manifest.json
│       └── skill-creator/                     ← Anthropic's official skill-creator
│                                                  bundled for when you add more skills
├── .env                                       ← ZERNIO_API_KEY (placeholder until you fill it)
├── CLAUDE.md                                  ← architectural map of the repo
├── README.md
├── LICENSE                                    ← MIT
└── .gitignore

Keep the folder structure intact. Claude needs reference/platforms/ to look up rules when writing captions.

Step 3: Add the skill to your project

Open VS Code or use Claude Code / Cowork directly
Create a new folder for this project or use an existing one
Extract the zip and copy the entire folder to your project root
Your file tree should show .claude/zernio-publishing/ at the top level

In Cowork: Go to Skills → + Create skill → Upload skill and select the folder.

Step 4: Set up your API key

Claude needs a Zernio API key to publish.

Log in to Zernio
Go to API keys (settings)
Click Create key, name it (e.g., "claude-publishing")
Copy it

Create a .env file in your project root:ZERNIO_API_KEY=your-key-here

Claude Code reads this automatically—never paste keys into prompts.

Step 5: Test with a simple prompt

Start with one video and 2–3 platforms to confirm the skill works.

Give Claude this:

I have a short-form video at [path/to/video.mp4]. Can you analyze it, create optimized posts for Instagram, TikTok, and LinkedIn, and schedule them using Zernio?

Claude will:

Load the skill and platform guidelines
Analyze your video
Generate captions, hashtags, titles, and first comments
Call Zernio to stage the posts
Show you a preview before publishing

Step 6: Review before you approve

Claude shows all posts before publishing. You control what goes live.

You'll see something like:

Request changes if needed. Say "Approve" when it looks good.

Note: First comments are easy to miss if you skim the preview. Read the full output before approving—Instagram, TikTok, and LinkedIn generate engagement hooks you might not see at first glance.

Step 7: Scale to multiple platforms

Once you're confident with a few, the system handles format conversion and batching automatically.

You can:

Post to 9 platforms at once: Video or carousel becomes optimized for each platform's native format
Convert formats: Carousels become videos for platforms that don't support multi-image
Batch multiple assets: Claude processes everything and schedules it all

Example:

I have a carousel at [Google Drive link] and a 30-second clip at [path]. Convert the carousel to video and post both to Instagram, TikTok, X, YouTube, LinkedIn, and Threads. Optimize each for that platform. Schedule for tomorrow at 9 AM.

What's next

You own a system that turns one prompt into nine scheduled posts—all optimized and formatted. A video that took 2 hours to post manually now takes 5 minutes to set up and approve.

The time saved compounds. This week: 2 hours back. Next month: 20+ hours back. That's time spent creating instead of grinding.

Check posts in Zernio or directly on each platform. Track analytics and iterate. The system learns as you refine your voice.

Questions or running into issues? Drop them in the comments.