GitHub: github.com/csr1010/wappy-kit
Personal AI agents are having a moment. Muse, Instinct, Dots. I've watched the demos and they're impressive. I'll probably use a few of them myself.
But watching them, one thought kept coming back to me: what if the agent was mine?
Not mine as in "my account on someone's platform." Mine as in running on my own machine. My conversations, my memory and my data in one file I can open and read. My tools, my prompts, my rules. If I want it to behave differently, I change it this afternoon.
And I talk to it from the app I already have open all day: WhatsApp.
The bigger picture
WhatsApp is already where billions of people talk to their family, their friends, their barber, their kid's school. It's the closest thing we have to a universal interface.
Now imagine anyone could put their own agent there. Not a generic chatbot, but an agent that knows their world and can act on it:
- Your second brain. Feed it your notes, docs, manuals and recipes. Ask "what was the Wi-Fi password at the cabin?" or "what did we decide about the kitchen remodel?" and get an answer grounded in your own stuff.
- Your inbox and calendar, conversational. "Anything urgent from school this week?" "Am I free Thursday afternoon?"
- A small business that never sleeps. A restaurant number that checks real reservations before saying a table is free. A bakery that takes custom cake orders. A store that answers "Where's my order?" by looking at the actual order system, not a canned reply.
- Support that knows the product. Load your help docs and let the agent answer from them, then hand off to a human when it should.
- A team ops bot. Check inventory, look up a runbook or file a ticket from the group chat.
The same pattern runs through all of them: WhatsApp as the interface, your systems as the source of truth, and you in control of the agent in between.
None of these are pre-built integrations. They're what becomes easy to build once the hard plumbing is out of the way.
Why this never happened before
The agent itself was never the hard part. Everything around it was: WhatsApp webhooks, message formatting, memory that survives across conversations, model calls, tool wiring. A bot that says hello takes an evening. One that can look things up, remember what you meant and reply usefully is a whole project before you even get to your own idea.
So I built Wappy
Wappy is a free, open-source (MIT) TypeScript SDK for building your own AI agent on WhatsApp. It handles the plumbing so you can focus on the part that makes your agent yours.
- Tool calling with real context, so your agent can act, not just chat
- Any model you like: Anthropic, OpenAI, Gemini, or Ollama if you want the model running locally too
- Memory you own: a local SQLite/LibSQL file by default, or a self-hosted Cognee knowledge graph
- Knowledge base ingestion with grounded retrieval, so answers come from your data
- No Wappy account and no Wappy-hosted backend
It ships with a few examples to learn from: a read-only Gmail and Calendar assistant, persistent personal memory, and knowledge base Q&A. Use them, extend them, or ignore them and build something completely different.
Try it
You need Node.js 22 or newer:
npm create @wappy_ai/agent
The generator asks a few questions and scaffolds your agent in about a minute. Going live on WhatsApp takes a little longer. The generated guide walks you through setting up your own Meta Cloud API app and webhook.
What "local" means here
Your agent, its memory and your data live on your machine. WhatsApp messages still pass through Meta, and if you choose a hosted model, it receives your prompts. Pick Ollama if you want the model on your machine as well.
Over to you
This is early, and I'd love to see where people take it. If you try it, tell me what was hard.
And if you had an agent on WhatsApp that ran on your own machine with your own data, what would you build first?
GitHub: github.com/csr1010/wappy-kit
Top comments (0)