Every time I used ChatGPT or similar tools the same thought crossed my mind:
"Where is this conversation going? Who has access to it? What are they doing with my data?"
So I decided to build my own solution. Meet CrustAI 🦀 — a fully private, self-hosted AI assistant that runs entirely on your own machine.
No cloud. No subscriptions. No data leaving your computer. Ever.
🌐 Documentation | ⭐ GitHub
The Problem I Wanted to Solve
Most AI assistants have one fundamental issue — your data belongs to someone else.
When you ask ChatGPT something personal, that conversation is:
- Stored on OpenAI's servers
- Potentially used to train future models
- Subject to their privacy policy (which can change)
- Accessible in case of a data breach
I wanted an AI assistant that was genuinely mine — one that I could use daily through apps I already use (Telegram, WhatsApp) without worrying about privacy.
What is CrustAI?
CrustAI is a self-hosted AI assistant powered by Ollama that connects to your favorite messaging platforms and runs completely offline after setup.
Key features:
- 🔒 100% Private — all processing happens on your hardware
- 🧠 Local LLM — powered by Ollama (llama3.2, tinyllama, phi3, mistral...)
- 📱 Multi-platform — Telegram, WhatsApp, Discord and Slack
- 🧬 Long-term memory — remembers facts across conversations
- 🗣️ Offline voice — speech-to-text and text-to-speech (pt-BR)
- ⚡ REST API — built-in Fastify server for custom integrations
- 🎭 Custom personality — fully configurable assistant behavior
Tech Stack
| Technology | Purpose |
|---|---|
| Node.js | Runtime environment |
| Ollama | Local LLM inference engine |
| node-telegram-bot-api | Telegram integration |
| @whiskeysockets/baileys | WhatsApp integration |
| discord.js | Discord integration |
| @slack/bolt | Slack integration |
| Fastify | REST API server |
| sql.js | Embedded SQLite for memory |
| yaml | Configuration management |
Architecture Overview
The architecture is straightforward — a central message handler receives messages from any platform and routes them through the same pipeline:
┌─────────────────────────────────────────────────────┐
│ CrustAI Core │
│ │
│ ┌──────────┐ ┌──────────┐ ┌──────────────────┐ │
│ │ Telegram │ │ Discord │ │ WhatsApp │ │
│ │ Adapter │ │ Adapter │ │ Adapter │ │
│ └────┬─────┘ └────┬─────┘ └────────┬─────────┘ │
│ └──────────────┼──────────────────┘ │
│ ▼ │
│ ┌─────────────────┐ │
│ │ Message Handler │ │
│ └────────┬────────┘ │
│ ┌───────────┼───────────┐ │
│ ▼ ▼ ▼ │
│ ┌─────────┐ ┌────────┐ ┌──────────┐ │
│ │ Ollama │ │ Memory │ │ REST API │ │
│ │ (LLM) │ │ Store │ │ Server │ │
│ └─────────┘ └────────┘ └──────────┘ │
└─────────────────────────────────────────────────────┘
Each platform adapter is independent — you enable only what you need in config.yml.
Getting Started in 5 Steps
Prerequisites
- Node.js ≥ 20
- Ollama installed
- A Telegram bot token from @BotFather
Step 1 — Clone the repository
git clone https://github.com/DaveSimoes/CrustAI.git
cd CrustAI
Step 2 — Install dependencies
npm install
Step 3 — Start Ollama and pull a model
# Start the Ollama server (keep this terminal open)
ollama serve
# In a new terminal — pull a lightweight model
ollama pull tinyllama # 600MB — works on modest hardware
# or
ollama pull llama3.2 # 2GB — needs ~8GB RAM
Step 4 — Configure the project
cp config/config.example.yml config/config.yml
Edit config/config.yml:
model: tinyllama
ollama_url: http://localhost:11434
language: pt-BR
telegram:
enabled: true
token: YOUR_BOT_TOKEN_HERE
allowed_user_ids: []
Step 5 — Launch CrustAI!
npm start
Expected output:
✓ Ollama connected (tinyllama)
✓ Memory store ready (./data/memory.db)
✓ Telegram ready
✓ REST API ready (http://localhost:3000)
🦀 CrustAI is ready. Your shell awaits.
Available Commands
Once running, use these commands directly in Telegram (or any connected platform):
| Command | Description |
|---|---|
/ping |
Check if the bot is alive |
/help |
Show all commands |
/model |
Show which AI model is running |
/remember <fact> |
Store a fact in long-term memory |
/forget |
Erase all stored facts |
/clear |
Clear conversation history |
The Privacy-First Philosophy
CrustAI was built with privacy as its core principle, not an afterthought:
✅ All conversations processed locally — nothing leaves your hardware
✅ No calls to OpenAI, Anthropic, Google or any cloud AI service
✅ No telemetry or usage tracking
✅ Open source — audit every single line of code
✅ MIT license — use it however you want
Running on Modest Hardware
One concern I had was performance. I'm running CrustAI with tinyllama on a machine with limited RAM and it handles daily conversations well.
For basic Q&A and conversation, tinyllama is surprisingly capable. If you have more resources, llama3.2 or phi3 give significantly better results.
What's Next — Roadmap
- [ ] Web UI dashboard
- [ ] Image understanding (multimodal LLMs)
- [ ] Plugin system for custom tools
- [ ] Docker one-click deployment
- [ ] Mobile app companion
Contribute!
CrustAI is open source and I'd love contributions from the community:
- 🐛 Bug reports — open an issue
- 💡 Feature ideas — let's discuss in issues
- 🔧 Pull requests — always welcome
- ⭐ Star the repo — helps others discover it!
GitHub: https://github.com/DaveSimoes/CrustAI
Documentation: https://documentcrustai.netlify.app
Final Thoughts
Building CrustAI taught me a lot about local LLM inference, multi-platform bot development, and the real meaning of privacy-first software.
If you're tired of sending your conversations to the cloud, give CrustAI a try. Your data deserves to stay yours.
Made with 🦀 and ❤️ by Dave Simoes
If you found this useful, consider starring the repo and sharing with others who care about privacy!