WhatsApp chatbot or your own AI? Connect the AI you already use to your WhatsApp Business number
Instead of building a chatbot inside a vendor's platform, you can connect the AI you already use, a Claude or GPT model or any model on OpenRouter, to your WhatsApp Business number through Meta's official Cloud API. Each incoming message goes to your model with your instructions and the recent chat, its reply goes back on the same chat, and with Coexistence you keep answering from the app.
easyCoexistence connects your WhatsApp Business number to Meta's Cloud API through Coexistence and runs your own AI on it, in a Reply with AI step built in the editor or by asking ChatGPT or Claude through the MCP connector.
Anthropic, OpenAI and OpenRouter keys work in the Reply with AI step, with any model id the key can use.
Listed in OpenRouter's public catalog on 11 November 2026, including Gemini, Llama, DeepSeek, Mistral and Qwen.
Default pause on a chat after the owner replies from the WhatsApp Business app, adjustable from 0 to 168.
Default cap on AI replies per chat, adjustable from 1 to 60, which cuts off two bots replying to each other.
Meta's policy allows automated replies in this window if customers get prompt, clear and direct paths to a human.
How does your own AI reply on a WhatsApp Business number?
Through an automation that reacts to incoming messages. A customer writes to your number, Meta delivers the message to easyCoexistence through the official Cloud API, and the automation whose trigger matches runs its steps. The "Reply with AI" step sends your Instructions and the latest messages of that chat to the model you picked, on your own key, then sends the model's answer back on the same chat. The answer also shows in the chat thread of the WhatsApp Business app, so whoever holds the phone sees what your AI said. No code is involved: start from scratch in the editor and add a "Reply with AI" step, pick "Agent answers everything" for an AI agent that also reads uploaded files and pages, or ask ChatGPT or Claude to build it through the MCP connector, which calls create_automation and then test_automation.
Chatbot builder or your own AI: what is the difference?
Who owns the AI. A chatbot builder sells you its bot, built in its editor with the AI it offers. Connecting your own AI keeps the model, the key and the bill yours, on the number you already use; easyCoexistence is the connection and the automations around it, not a bot.
Which AI models can answer customers?
Any model your key reaches at three providers: Anthropic for Claude models, OpenAI for GPT models, and OpenRouter, where one key reaches models from many labs. Save the key once with the "AI keys" button on the Automations page. "Test and save" checks it with the provider without spending tokens and loads the model list; a refused key is not saved. Afterwards the key shows only as "ending in 1234", stored encrypted and used only for your automations' AI replies. Keys are entered in the dashboard, never through a chat or the MCP connector. Each Reply with AI step names its own provider and model, so a keyword automation can run a small model while the catch-all runs a larger one.
What does the AI read before it answers?
Your Instructions, up to 20,000 characters, and the chat's earlier messages, set by "Earlier messages it reads" from 0 to 20, with 10 by default. Those messages include the customer's, earlier automated replies and the replies you typed in the app. With automations on, message text is kept for the period you choose in Conversations, 30, 90 or 365 days (90 by default), and the model reads up to 20 of those earlier messages. Run logs keep the message that started each run and the reply for 30 days. Under your Instructions the runner adds five short rules on every AI reply. The Reply with AI step reads text and answers with text only: button and list replies arrive as the text of the option tapped, and media arrive as placeholders (an AI agent can also look at pictures and PDFs).
- Photos, videos and documents: "[image]", "[video]" or "[document]", followed by the caption when there is one.
- Voice notes: "[voice message]" followed by the words when an OpenAI or OpenRouter key is saved, which writes them out; without one, just "[voice message]".
- Stickers, locations and contacts: "[sticker]", "[location]" with the place name, and "[contact card]".
- The added rules: a short plain-text message in the customer's language, a person offered when asked, no invented prices or policies, and an honest answer that it is automated. The first AI reply says so itself.
What happens when you answer a chat yourself?
Your AI steps back on that chat. A message you type in the WhatsApp Business app reaches easyCoexistence as an echo, and the automation stays quiet on that chat for the hours set in "Pause on a chat after I answer it (hours)": 4 by default, anywhere from 0 to 168. A follow-up still waiting on that chat is dropped as well, while other chats keep their automated replies. When a customer asks for a person, the added rules already make the model say someone from the business will answer soon, and your reply from the phone takes over with no switch to flip. The customer sees one conversation with one business.
Which limits keep AI replies under control?
The runner enforces these by itself, whatever the Instructions say, and the run log names the one that stopped a reply:
- One automation per message: keyword rules answer first, then First message, then After hours, and the catch-all last.
- Text and AI replies go out only within 24 hours of the customer's last message; after that, only an approved template.
- "AI replies per chat per hour, at most" is 10 by default (1 to 60), so two bots answering each other stop quickly.
- "Longest reply (tokens)" caps an answer at 600 tokens by default, adjustable from 50 to 2,000.
- A message Meta delivers twice gets one reply, and reactions, statuses and your own sends never trigger one.
- Sends are spaced to at most 10 per second per number, half the 20 per second Meta fixes for Coexistence numbers.
- A failed reply emails you, at most once per automation every 6 hours, so you can answer that customer yourself.
What does it cost to run your own AI on WhatsApp?
Three separate bills, and only one comes from us. easyCoexistence charges per connected number, with automations included: no cap on runs and no charge per AI reply. Your AI provider bills the tokens to your own key at its own rates. A reply uses your Instructions, the earlier messages it reads and the answer itself, and tests in "Try it" spend real tokens too. Meta bills your own WhatsApp Business Account: every number has a free tier of 1,000 service messages a month, Meta charges from the 1,001st, and templates are billed per message by category and country; we never mark them up. Trimming the Instructions and lowering "Earlier messages it reads" cut the AI bill more than anything else.
How to set it up
- Connect your numberConnect your WhatsApp Business number through Coexistence by scanning the code from the app; nothing changes on the phone. New connections start with "Automations on". A number set to "Direct to your server" needs "Turn on automations" on its page first.
- Save your AI keyOn the Automations tab, open "AI keys", choose Anthropic, OpenAI or OpenRouter, paste the API key and press "Test and save".
- Start from the AI assistant examplePress "New automation", pick "AI assistant" and press "Use this". It answers whatever the customer writes, with one "Reply with AI" step.
- Write the Instructions and pick a modelDescribe the business, the facts the AI may use and when to hand over to a person, then pick a model from the list your key loaded.
- Test it before customers see itSwitch it Off, press Save, open Test and use "Run test" with real questions. Nothing is sent, but your model really answers. Switch it On and save when the replies read right.
Common mistakes
- Saving a new automation without switching it Off. It saves On by default, so it starts answering real customers before you have run a single test.
- Reply with AI gets "[image]" plus any caption, not photos, and voice notes as "[voice message]" with words only if an OpenAI or OpenRouter key is saved; tell it in Instructions to ask customers to type. An AI agent reads pictures and PDFs.
- Expecting the AI to answer a message a keyword automation already took. Only one automation answers, and keyword rules win, so add a Reply with AI step inside the keyword automation when both should happen.
- Letting the AI answer a burst of short messages one by one. Three quick lines get three replies; a 1-minute Wait before Reply with AI lets each new line replace the waiting run, so one reply covers all three.
Automations cost nothing extra, and your AI provider bills its tokens to you with no markup from us. US$ 9 down to US$ 0.75 per number per month, falling as you add numbers, with the first 3 days free.
easyCoexistence is a multichannel API and MCP server that connects AI agents such as Claude and ChatGPT to WhatsApp, Instagram, TikTok, Facebook and more. On WhatsApp it uses Coexistence: the business keeps using its number in the WhatsApp Business app while the AI works through Meta's official API.
Frequently asked questions
Can I use my ChatGPT or Claude subscription instead of an API key?
No. The Reply with AI step needs an API key from the provider's developer platform, where usage is billed per token. A chat subscription is a separate product: OpenAI and Anthropic bill API usage on their developer platforms, apart from ChatGPT and Claude plans.
Is a business allowed to answer WhatsApp customers with AI?
Yes, within Meta's rules. Section 4.7 of Meta's WhatsApp Business terms bars providers of general-purpose AI assistants when the assistant is the primary functionality offered, rather than incidental or ancillary, as Meta determines; it is not aimed at a business using AI to serve its own customers. Meta's messaging policy also asks automated replies to offer a clear path to a person.
Will my AI answer in Portuguese or Spanish?
Yes. The rules added under your Instructions tell the model to reply in the customer's language, so one assistant can answer in Portuguese, Spanish and English. Write the Instructions in the language you write most precisely, and include the local terms customers use, such as Pix or CEP.
What happens if my AI key stops working?
If the provider refuses the key (HTTP 401 or 403, or 402 for OpenRouter credit), every automation using that provider shows "Paused: AI key refused" and you get one email saying the key stopped working. While paused they are left out of matching, so a lower-priority automation can answer instead. Saving a working key under AI keys restarts them. Other provider errors fail only that run, and you are emailed about that customer.
Keep reading
Ready to get started?
Set up WhatsApp Coexistence in minutes, not months. The app keeps working on the phone.
Start Free TrialVerified on