Hello,
I’m looking for a way to create and publish a free, 24/7 AI-powered WhatsApp bot/agent.
I need it to:
Connect to WhatsApp
Use AI to respond automatically
Be active 24/7
Be completely free of charge
If anyone knows of any tools, platforms, or methods to achieve this, I would greatly appreciate the help.
So what you’re actually asking — let me say it back:
“How do I make a WhatsApp number that auto-replies with AI to anyone who texts it, runs on its own forever, and costs me nothing.”
Your 4 stated needs:
That last one kills 87% of new bots, and every consensus tutorial silently leaves you to discover it the hard way. We’ll handle it.
Sounds like you want this to actually survive on its own, not just exist for a week — so I’m skipping the Twilio + OpenAI-key answer that’s been the standard reply for 5 years and is genuinely outdated now.
The dumb-proof timeline
Right now (5 min) → pick your shape of bot from the table below. This one decision changes everything else.
This weekend (2–3 hrs) → fork lyfe00011/levanter → one-click deploy to Koyeb → plug in a free Cerebras key
Across the next 7 days → run the warming protocol before your number does any AI work
Total cost → $0/month, no credit card needed at any step if you choose right
Pick your shape first — this saves you days| What you want the bot to do | Best path | Cost | Ban risk | Time |
|---|---|---|---|---|
| Auto-reply when people DM me (support / FAQ / “I’m away”) |  Cloud API |
$0 forever | None | 1 hr |
| Be my personal AI on WhatsApp (like ChatGPT but on my number) |  Baileys + Levanter |
$0 + a phone | Real, manageable | 3 hrs + warm |
| Group bot for friends (commands, AI, stickers) | Levanter on Termux |
$0 | Low (groups are friendlier) | 2 hrs |
| Talk to my AI from WhatsApp without a WA-side bot |  Telegram ↔ WA bridge |
$0 | None | 4 hrs |
Two things consensus tutorials still get wrongReveal #1 — The official Cloud API has been free AND unlimited for incoming-conversation replies since November 1, 2024. Meta quietly removed the 1,000-conversation cap. If your bot only replies to people who messaged you first (within the 24-hour reply window), the official path is free, ban-immune, and easier than every unofficial alternative. Most “free WhatsApp bot” guides still tell you the API costs money. It doesn’t.
Reveal #2 — WhatsApp banned ChatGPT-style “ask anything” bots from the Cloud API on Jan 15, 2026 (TechCrunch). Task-specific bots are fine: support, FAQ, booking, study buddy, recipe helper. Open-ended assistants need to live on Path B (personal account via Baileys), which means the warming protocol becomes non-negotiable.
The “I use this” — what’s actually working in my drawerI run my own setup on a Redmi 6A I had lying around — Termux + Levanter + Ollama with phi3:mini for the AI part. Total monthly cost is the electricity to keep the charger plugged in.
Got bit on day 4 once because I joined two groups too quickly and earned a 24-hour restriction. Now I add max one group per day during week 1. Phone’s been online ~11 months without a permanent ban.
That’s the “I use this” — not a flex, just what survived for me.
↓ EXPAND THE PATH YOU PICKED ↓
Best if your use case is “auto-reply when someone DMs my number.” Officially supported, free forever for incoming conversations, and as of Nov 2024 it has no message cap.
1. Hit Meta for Developers → Cloud API quickstart. Click Get started.
What happens: it walks you through making a “Meta Business” — basically a placeholder business profile, not a registered company. You see green checkmarks as each step completes.
2. Get your test phone number. Meta hands you a free test number plus 5 recipient slots, no card needed.
What you’ll see: a number like
+1 555 xxxxin the dashboard. That’s your bot’s number for testing.
3. Wire it to a free LLM. Get a Cerebras free key (1M tokens/day, no card, instant). Glue them together using n8n’s WhatsApp Cloud API node — n8n’s free cloud tier handles the webhook side without you running any server.
4. When you’re ready for real users, request your own number through Meta.
What happens: verification takes ~1 day. Fresh numbers start at 250 unique conversations / 24h, automatically scale to 1,000 → 10,000 → unlimited as your “quality rating” stays green.
The framing trick that keeps you policy-compliant: describe your bot as “FAQ assistant” or “support bot” on the registration form, not “AI chatbot.” Meta banned general-purpose AI in Jan 2026, but “AI-powered support” is explicitly fine. Same code, different label.
Best if you want the bot on a personal-style number, no policy restrictions, full control. Trade-off: you accept the ban risk and run the warming protocol.
Three options, ranked by my honest preference:
Old Android phone via Termux — install Termux from F-Droid only. The Play Store version is outdated and missing newer packages. Truly free, runs forever, survives reboots.
Oracle Cloud Always Free — 4 ARM cores, 24GB RAM, permanent. Card needed at signup but never charged.
Koyeb free tier — simplest. Levanter has a 1-click deploy button right in the README.
Don’t use Render’s free tier for this specific job. Render sleeps the service after 15 minutes idle, which kills the WebSocket connection Baileys relies on. You’ll spend hours wondering why the bot keeps dying. Other free hosts work fine — just not Render here.
lyfe00011/levanter is a 2k-star, multi-session WhatsApp bot framework built on Baileys. It’s what you’d build after 3 weeks of work, sitting there ready to fork.
What you’ll see: in the README, deploy buttons for Koyeb, Render, Heroku, and a Panel option. Click your host, fill in the env vars, done.
git clone https://github.com/lyfe00011/levanter botName
cd botName
yarn install --network-concurrency 1
Then visit Levanter’s pairing flow, enter your number, get an 8-digit code, type it into WhatsApp on your phone:
Settings → Linked Devices → Link with phone number
How you know it worked: your phone shows “Linked successfully” and the bot terminal stops printing connection-retry messages.
Levanter has a plugins/ folder. Drop in a JS file that hits your chosen free LLM. See the 
Free LLM picker section below for which provider fits your workload.
Open the Day-by-day warming section below. This is the part most guides leave out, and the reason most people fail.
Best if your real intent is “I want to chat with my AI from WhatsApp” and you don’t need to expose the bot publicly to other people.
Your bot lives on Telegram (where bots are first-class & free, never banned). Your personal WhatsApp account just bridges to it through a Matrix homeserver. You message the bot from WhatsApp → bridge routes it to Telegram → bot replies → bridge sends it back to WhatsApp. The bot side never touches WhatsApp’s anti-bot infrastructure.
Easiest: Beeper (free, hosted, just sign up and link)
Self-hosted: Conduit homeserver (Rust, runs on a $5 VPS or Pi) + mautrix-whatsapp bridge
self_sign: truein mautrix-whatsapp’s config. Otherwise the bridge breaks in April 2026 (per the maintainers’ troubleshooting note). This took out a lot of bridges quietly for people who didn’t know.
This path was completely missing from every English-language WhatsApp-bot tutorial I checked while researching this. Often the cleanest fit if your real goal is just “I want my AI on WhatsApp.”
Best if you want zero API keys anywhere, ever. Lower model quality but no rate limits and no provider can shut you down.
# In Termux on the same phone running Levanter:
pkg install ollama
ollama serve &
ollama pull phi3:mini # 2.3GB — works on most phones
# OR
ollama pull gemma2:2b # 1.6GB — fastest on mobile
Point Levanter at http://localhost:11434/v1/chat/completions — Ollama exposes an OpenAI-compatible endpoint, so most plugins just work with the URL swapped.
Confirmed working: step-by-step walkthrough on dev.to.
The trick: WhatsApp’s anti-spam doesn’t read your messages. It reads your behavior. A new number sending 50 messages in the first hour looks like a bot. The same number behaving like a slow human for a week earns a “trusted” tag and stays alive for years.
Source: GREEN-API’s published anti-ban routine, validated against WhatSnap’s analysis showing 87% of new numbers get banned in the first 72 hours without it.
| Day | What to do | What NOT to do |
|---|---|---|
| 0 | Verify number on a real phone, real SIM. Set human profile photo (NOT a logo). Use a first name. Open WhatsApp 20–30 min. | Don’t link Web/Desktop yet. Don’t send any message. |
| 1 | Wait 24h after registration before pairing. Receive 1–2 messages from a friend. Reply 2–4 hours later. | Don’t initiate. Don’t join groups. |
| 2–4 | Receive 1 message every ~2h from people who already saved your number. React with emojis. Update profile slowly. Join 1 small group (<15 people), lurk only. | Don’t message strangers. Don’t send identical templates. |
| 4 | Bot starts replying — same human pace, only to contacts who have you saved. | Don’t broadcast. Don’t reply with identical timing each time. |
| 5–7 | Scale from 12 → 100 messages/day. Vary timing. Vary length. Personalize. | No status uploads from the bot — this triggers permanent bans, see Baileys Issue #2309. |
| 10 | Materially ban-resistant.  |
|
| 25–30 | Algorithm gives “green light” — stable. |
| Provider | Best for | Daily limit | Speed | Direct signup |
|---|---|---|---|---|
| Cerebras | Volume (highest free quota) | 1M tokens/day | ~2,600 tok/sec | cloud.cerebras.ai |
| Groq Llama 3.1 8B | Speed (fastest) | 14,400 requests | ~1,800 tok/sec | console.groq.com |
| Gemini 2.5 Flash-Lite | Long context (1M token window) | 1,000 requests | medium | aistudio.google.com |
| OpenRouter | Variety (rotate free models) | varies by model | varies | openrouter.ai |
| Local (Ollama) | Zero dependency | ∞ | hardware-bound | ollama.com |
Verified rate limits as of April 2026. aifreeapi keeps a current table.
Speed-critical chatbot → Groq Llama 3.1 8B
Volume-critical → Cerebras (1M tokens/day is more than any single-user bot will use)
Long documents / context → Gemini 2.5 Flash-Lite
Privacy / offline → Phi-3 Mini local on Termux
Closing the loopYou said “create and publish” — that “publish” word is what made me think you want this to actually run reliably for real people, not just be a weekend toy. So one question back, because the answer changes which path I’d push you toward first:
When you said “completely free of charge” — were you thinking “no credit card anywhere, at any step” or “no recurring monthly bill”?
Oracle’s permanent free tier needs a card at signup but never charges. The Termux-on-old-phone path needs zero anything. The first opens up better hosting; the second is bulletproof against ever being charged. Knowing which one you meant pins down the exact recommendation.
And to anyone lurking who’s run Levanter or any Baileys bot for 3+ months without a ban — what did your warming routine actually look like in practice? The published protocols are good, but I suspect the day-1-to-day-10 reality is messier and the polished version skips the parts that actually mattered. Real numbers > polished ones.
!