cd /news/ai-agents/building-an-agent-that-reads-whatsap… · home › topics › ai-agents › article
[ARTICLE · art-172602] src=dev.to ↗ pub= topic=ai-agents verified=true sentiment=↑ positive

Building an agent that reads WhatsApp notifications and never sends on its own

Itamar, a construction business owner, built Focus 100, an Android app that uses a NotificationListenerService limited to user-selected apps to read incoming WhatsApp and email notifications and convert them into structured tasks via Claude. The system sends each message to a stateless server that returns structured JSON (task, promise, event, or noise), then applies deterministic code for date fixes, group rules, deduplication, and calendar conflicts, with every outgoing message requiring an editable card and manual Send button. The developer reports that a self-hosted Gemma model matched Claude Haiku on a small eval but suffered 110-170 second cold starts, prompting a switch to Claude at roughly $0.0006 per request with the open model kept as fallback.

recommendation

Keep the LLM stateless and validate its structured JSON with deterministic code for dates, group rules, dedup and conflicts, and gate every outgoing message behind an editable card with a manual Send button.

more in playbook: Designing agent systems
read1 min views5 publishedNov 19, 2026

I'm Itamar. I run a construction business, and most of my real tasks arrive in WhatsApp groups, voice notes and email. I built Focus 100 to read those and turn them into tasks. Here's what mattered technically.

On Android, a NotificationListenerService limited to the apps the user picks (WhatsApp, WhatsApp Business, mail apps). It only sees new messages that post a notification, so muted chats are invisible. That limit is a feature for privacy: no scraping, no linked device.

Each message (with a bit of same-chat context) goes to a stateless server that asks Claude for structured JSON: task / promise-to-me / event / noise, title, due date, suggested reply. Then plain code checks it: weekday and "tomorrow" fixes, group rules (a task only when you're named, tagged, or everyone is asked), dedup, calendar clashes.

Poki, the assistant, receives a contact summary (names, relations, which channels exist, the usual channel) but no numbers or addresses. The phone fills those in at execution time. Every outgoing message is a card with editable text and a Send button.

Our own open model (Gemma on a serverless GPU) was about as accurate as Claude Haiku on a small eval, but cold starts were 110-170 s. We switched to Claude (~$0.0006/request) and kept the open model as a fallback.

Tasks live on the device. Optional sync is end-to-end encrypted (AES-GCM, key only on devices); the server holds one ciphertext per space; merge is last-write-wins per task with tombstones.

iOS doesn't let apps read other apps' notifications, so on iPhone it's a web app (paste, import an exported chat, or talk to Poki).

It's free, no account, no ads; web app in 15 languages. I'm the developer, and I'd love feedback on false positives in busy group chats.

Android: [https://focus100-il.netlify.app/get?ref=devto](https://focus100-il.netlify.app/get?ref=devto)

Web: [https://focus100-ios.netlify.app](https://focus100-ios.netlify.app)
── more in #ai-agents 4 stories · sorted by recency
── more on @focus 100 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/building-an-agent-th…] indexed:0 read:1min 2026-11-19 · —