agi.diy

Build your own AGI. In your browser. Right now.

▶️ Launch • 📖 SDK Docs • 📱 Install as App

---

Why agi.diy?

	ChatGPT/Claude.ai	agi.diy
Privacy	Data on their servers	100% in your browser
Cost	$20-200/month subscription	Pay only for API usage
Offline Mode	❌	✅ WebLLM runs locally
Custom Tools	Limited plugins	Create unlimited tools
Multi-Agent	❌	✅ Coordinated agent teams
Self-Modifying	❌	✅ Agent evolves itself
Open Source	❌	✅ Fully auditable

---

Quick Start

Use the hosted version:

open https://agi.diy

Or self-host:

git clone https://github.com/cagataycali/agi-diy.git
cd agi-diy/docs && python3 -m http.server 8080

Then: Settings → Add API key → Start chatting

Want 100% local? Select WebLLM → Download model once → Works offline forever

---

How It Works

graph LR
    A[You type message] --> B[Agent selects tools]
    B --> C[Calls AI model]
    C --> D[Executes in browser]
    D --> E[Shows result]
    
    style A fill:#3b82f6,color:#fff
    style E fill:#10b981,color:#fff

Loading

Everything runs in your browser. The only external call is to the AI provider you choose (or none with WebLLM).

---

What Can It Do?

🛠️ Create Custom Tools

Ask the agent to build tools on the fly:

"Create a tool that fetches Bitcoin price from CoinGecko"

The tool saves to localStorage and persists forever. Use it anytime:

"What's the current Bitcoin price?"
→ Uses your custom tool → "$67,432"

📧 Automate Email

"Every morning at 9am, check my Gmail and notify me of urgent emails"

Agent connects via Google OAuth, schedules a cron job, and sends push notifications.

💻 Pair Programming

"Watch my screen every 30 seconds and help me debug"

Agent captures your screen, spots errors, and suggests fixes in real-time.

👥 Multi-Agent Research

"Spawn researcher, analyst, and writer agents. Research AI safety."

Agents coordinate through ring attention—when one learns something, others see it.

🌙 Background Thinking

Ask about a topic, then walk away. Agent keeps exploring while you're gone. When you return, findings are injected into your next message.

🗺️ Location Intelligence

"Mark top 5 coffee shops near me and fly me through them"

Interactive Google Maps with GPS tracking and smooth camera animations.

---

Two Modes

Mode	File	What it's for
Single Agent	index.html	Personal assistant, coding help, quick tasks
Multi-Agent	agi.html	Research teams, parallel processing, scheduled automation

Multi-Agent Architecture

graph TD
    You[You] --> Main[Main Agent]
    Main --> R[Researcher]
    Main --> A[Analyst]  
    Main --> W[Writer]
    
    R -.-> Ring[(Ring Buffer)]
    A -.-> Ring
    W -.-> Ring
    Ring -.-> R
    Ring -.-> A
    Ring -.-> W

Loading

Ring Attention: Agents share context automatically. When the researcher finds papers, the analyst and writer see that context immediately.

---

Models

Cloud Models (API key required)

Provider	Models	Best for
Anthropic	Claude Opus, Sonnet, Haiku	Quality reasoning
OpenAI	GPT-4o, GPT-4, GPT-3.5	General tasks
Amazon Bedrock	Claude + extended thinking	Deep analysis

Local Models (free, offline)

Model	Size	Notes
Qwen 2.5 3B ⭐	~2GB	Recommended for most users
Qwen 2.5 1.5B	~1GB	Faster, less capable
Llama 3.2 1B	~700MB	Smallest, for mobile
Hermes 8B	~4GB	Best tool usage

WebLLM requires Chrome/Edge 113+ with WebGPU

---

Tools Reference

Core

• render_ui — Dynamic HTML components in chat
• javascript_eval — Execute JS, return results
• storage_get/set — Persistent localStorage
• fetch_url — HTTP requests
• notify — Push notifications (works in background)

Self-Modification

• create_tool — Define new tools at runtime
• list_tools — See all available tools
• delete_tool — Remove tools
• update_self — Rewrite system prompt

Vision & Context

• get_user_context — Activity state, mouse position, idle time
• set_context — Add custom context
• scan_bluetooth — Find nearby devices and agents

Maps

• add_map_marker — Place markers with emoji/labels
• fly_to_location — Smooth camera animations
• tour_markers — Animated journey through points
• get_map_location — Current GPS position

Google APIs

• google_auth — OAuth 2.0 authentication
• use_google — Access 200+ Google services
• gmail_send — Send emails directly

Multi-Agent (agi.html only)

• use_agent — Spawn sub-agents
• invoke_agent — Call agent, wait for response
• broadcast_to_agents — Message all agents
• scheduler — Cron-based recurring tasks

---

Ambient Mode

Agent thinks while you're away.

Mode	Trigger	Behavior
🌙 Standard	30s idle	Runs 3 iterations, then pauses
🚀 Autonomous	Click button	Runs until [AMBIENT_DONE] or stopped

How it works: You ask about quantum computing, go make coffee. Agent explores applications, recent breakthroughs, industry adoption. When you return, those findings auto-inject into your next message.

---

Privacy & Security

Your data never leaves your browser (except queries to your chosen AI provider).

What	Where it's stored
API Keys	localStorage (never transmitted)
Conversations	localStorage
Custom Tools	localStorage
Settings Sync	AES-256-GCM encrypted

With WebLLM: Zero external calls. Everything runs on your GPU.

---

Install as PWA

Platform	Steps
iOS	Safari → Share → Add to Home Screen
Android	Chrome → Menu → Install app
Desktop	Click install icon in URL bar

Features: Home screen icon, background notifications, offline support, settings sync between devices.

Sync Settings

1. Settings → Sync → Enter password
2. Copy encrypted URL
3. On other device: paste URL, enter password
4. All settings transfer securely

---

Configuration

Get API Keys

• Anthropic: console.anthropic.com
• OpenAI: platform.openai.com
• Bedrock: AWS Console → API Keys

Extended Thinking (Bedrock)

Paste in Settings → API → Additional Request Fields:

{
  "thinking": { "type": "enabled", "budget_tokens": 10000 }
}

Google OAuth

1. Cloud Console → Create OAuth Client
2. Add authorized origin: https://agi.diy
3. Settings → Google → Paste Client ID

URL Shortcuts

https://agi.diy/?q=what+time+is+it

Great for iOS Shortcuts—one tap to query.

---

Console API

agi.agent              // Agent instance
agi.clear()            // Clear conversation
agi.tools.list()       // List custom tools
agi.tools.delete(name) // Remove tool

agiContext.getContext()      // All context data
agiContext.scanBluetooth()   // Find nearby devices

---

Troubleshooting

Issue	Solution
No response	Settings → Check API key
WebLLM won't load	Use Chrome/Edge 113+
Model download stuck	Refresh page
Screen capture denied	Allow browser permission
No notifications	Enable in browser settings

---

Project Structure

docs/
├── index.html      # Single agent mode
├── agi.html        # Multi-agent mode
├── strands.js      # Strands SDK bundle
├── vision.js       # Screen capture, ambient mode
├── webllm.js       # Local model inference
├── map.js          # Google Maps integration
├── tools/google.js # Google API tools
├── sw.js           # Service worker (PWA)
└── manifest.json   # PWA config

---

Contributing

PRs welcome for:

• ✅ New tools
• ✅ Model providers
• ✅ Bug fixes
• ❌ Build systems
• ❌ Framework dependencies

The project is intentionally minimal—single HTML files, no build step, fully auditable.

---

License

Apache 2.0

---

Built with Strands Agents SDK
agi.diy