FeelyAI

Run advanced AI models directly in your browser. Private, secure, and purely local.

FeelyAI is a powerful, privacy-focused AI chat application that runs entirely in your browser using WebLLM. No server costs, no data sent to external servers, and 100% client-side execution.

🌐 Try Live Demo

Experience FeelyAI in action at feelyai.com

✨ Features

🔒 100% Private & Local

• All AI processing happens directly in your browser using WebGPU
• No data sent to external servers
• No API keys required
• Complete privacy and security

🤖 Multiple AI Models

Choose from several optimized language models:

• Llama 3.2 3B - Fast and efficient general-purpose model
• Hermes 3 (Llama 3.1 8B) - Advanced reasoning and tool use
• TinyLlama 1.1B - Ultra-lightweight for quick responses
• Qwen2.5-Coder-3B - Specialized for coding tasks

🛠️ MCP (Model Context Protocol) Integration

• Connect to MCP servers to extend AI capabilities
• Add custom tools and functions
• Support for custom headers and authentication
• Enable/disable servers on the fly
• Real-time tool discovery and schema introspection

🔧 Built-in Tools

FeelyAI includes powerful built-in tools:

• evalCode - Execute JavaScript code directly in the browser (with safe eval mode)
• listTools - Discover available tools with search and regex filtering
• getToolSchema - Inspect tool schemas before use

🎯 Smart Tool Execution

• Schema Discovery Pattern - AI automatically discovers and learns tool schemas
• Validation & Error Handling - Automatic validation with helpful error messages
• Permission System - Granular control over tool execution
  • Allow once
  • Allow for session
  • Always allow
• Safe Eval Mode - Run code in isolated Web Workers for security

💬 Advanced Chat Management

• Projects & Chats - Organize conversations into projects
• 🔐 Encrypted Projects - Password-protect sensitive projects with client-side AES-GCM encryption
  • All chats, messages, and system prompts encrypted locally
  • Auto-lock on page reload for security
  • Passwords hashed with bcryptjs (never stored in plain text)
  • Unique encryption keys per project using PBKDF2
• Persistent Storage - All chats saved locally using IndexedDB
• Auto-titling - Chats automatically named from first message
• Message History - Full conversation history with tool calls and outputs
• Markdown Support - Rich text rendering with syntax highlighting

⚙️ Customizable System Prompts

• Per-Project System Prompts - Each project can have its own custom system instructions
• Edit system instructions to customize AI behavior
• Template variables for dynamic tool injection ({{listTools}}, {{tool_names}})
• Reset to defaults anytime
• Persistent across sessions
• Encrypted along with project data for password-protected projects

🎨 Modern UI/UX

• Clean, dark-themed interface
• Real-time progress indicators during model loading
• Collapsible tool outputs for better readability
• Responsive design
• Smooth animations and transitions

🔄 Model Switching

• Switch between models on the fly
• Automatic model download and caching
• Progress tracking during model initialization

🚀 Getting Started

Prerequisites

• Modern browser with WebGPU support (Chrome 113+, Edge 113+)
• Sufficient RAM (4GB+ recommended)
• GPU with WebGPU support

Installation

1. Clone the repository:

git clone https://github.com/hasmcp/feelyai.git
cd feelyai

2. Install dependencies:

npm install

3. Run the development server:

npm run dev

4. Open your browser and navigate to http://localhost:5173

Building for Production

npm run build
npm run preview

🔌 Adding MCP Servers

1. Click the "Add Server" button in the sidebar
2. Enter the MCP server URL (e.g., http://localhost:3000/mcp)
3. Optionally add a name and custom headers
4. Click "Add Server"

The AI will automatically discover available tools from connected servers.

🛡️ Security Features

Safe Eval Mode

When enabled (recommended), JavaScript code execution happens in an isolated Web Worker, preventing access to:

• DOM manipulation
• Local storage
• Cookies
• Parent window context

Permission System

Control which tools the AI can execute:

• Allow Once - Single execution approval
• Allow for Session - Approve for current session
• Always Allow - Trust tool permanently

📦 Tech Stack

• Frontend Framework: Vue 3 (Composition API)
• Build Tool: Vite
• AI Engine: @mlc-ai/web-llm
• MCP Client: @modelcontextprotocol/sdk
• Styling: TailwindCSS 4
• Icons: Lucide Vue
• Markdown: markdown-it
• Storage: IndexedDB (via idb)
• Validation: Ajv (JSON Schema)
• Encryption: bcryptjs + Web Crypto API (AES-GCM)

🏗️ Architecture

src/
├── components/
│   ├── AddServerModal.vue    # MCP server configuration
│   ├── ProjectSidebar.vue     # Project/chat navigation
│   └── McpIcon.vue            # MCP branding
├── composables/
│   └── useChat.js             # Main chat logic & state
├── services/
│   ├── McpClient.js           # MCP protocol client
│   ├── MessageStore.js        # IndexedDB persistence
│   └── EncryptionService.js   # Client-side encryption
├── workers/
│   ├── llm.worker.js          # WebLLM worker
│   └── eval.worker.js         # Safe code execution
└── App.vue                    # Main application

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Created by

HasMCP - No-code/low-code MCP server framework with built-in auth, realtime logs and telemetry

📄 License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

🙏 Acknowledgments

• WebLLM - Browser-based LLM inference
• Model Context Protocol - Tool integration standard
• MLC AI - Machine Learning Compilation

🔗 Links

• WebLLM Documentation
• MCP Specification
• Report Issues

---

⚠️ Disclaimer

Usage of LLM models is subject to their respective license agreements. Please review the license terms for each model you use. FeelyAI is provided as-is, and AI-generated content may contain errors or inaccuracies.

---

Powered by WebLLM • No Server Costs • 100% Client-Side