Ember Browser

AI-Native Chromium Browser with Embedded Local Intelligence

Browsing at the speed of thought.

---

🚀 Vision

Ember is not just another browser; it is an intelligent agent that navigates the web alongside you. By embedding a fine-tuned Large Language Model (LLM) directly into the browser core, Ember transforms passive browsing into active delegation.

Users speak or type commands—"Find the best price for a humidifier on Coupang", "Summarize this email thread", "Book a meeting for tomorrow at 2 PM"—and Ember autonomously navigates, extracts, and reports.

Core Philosophy:

• Local-First Intelligence: 4B parameter model runs on-device for sub-100ms latency and privacy.
• Zen Aesthetics: Minimalist UI with vertical tabs and translucent panels to keep focus on content.
• EmberScript ecosystem: A shareable, secure automation protocol that bridges natural language and DOM actions.

---

✨ Key Capabilities

🧠 Embedded Local Intelligence

• Zero-Latency Execution: Powered by a fine-tuned Qwen3-4B model running via llama.cpp directly within the browser process.
• Privacy by Design: Raw page content and personal data stay local. Only high-level intent goes to the cloud when necessary.
• Multilingual Native: Seamlessly handles Korean (target market P0) and English commands.

🤖 Autonomous Agent (Side Panel)

• Natural Language Control: command the browser using voice or text.
• Ember Execution Protocol (EEP): A standardized JSON protocol that translates intent into safe DOM actions (CLICK, TYPE, EXTRACT, VERTIFY).
• Smart Fallback: Triage system routes complex reasoning to cloud models (Claude/GPT-4o) while keeping execution local.

⚡ EmberScript

• Programmable Web: A novel scripting system combining structured selectors with natural language prompts.
• Marketplace: Discover and install automation workflows for your favorite sites.
• Secure Sandbox: Strict permission model prevents unauthorized actions.

🎨 Zen-Inspired UI

• Vertical Tabs: Optimized for modern widescreen displays.
• Floating URL Bar: Disappears when not needed to maximize viewport.
• Adaptive Theming: "Ember Dark" default with warm amber accents; follows system preferences.

---

🏗️ Architecture

Ember is built as a soft fork of Chromium. We do not maintain a divergent source tree; instead, we apply a strict series of patches to specific Chromium release tags.

Repository Structure

ember/
├── patches/                 # Chromium modifications (The Source of Truth)
│   ├── core/                # Branding, URL schemes (ember://)
│   ├── ui/                  # Vertical tabs, side panel
│   ├── ai/                  # Llama.cpp hooks
│   └── emberscript/         # Scripting engine integration
├── src/ember/               # Ember-specific source code
│   ├── ai/                  # Orchestrator, Cloud Bridge, Local Engine
│   ├── ui/                  # React/WebUI components for Side Panel & Settings
│   └── emberscript/         # Runtime, Parser, and Marketplace
├── third_party/llama_cpp/   # Vendored llama.cpp (Submodule pinned to specific commit)
├── models/                  # Local LLM artifacts (downloaded on bootstrap)
└── scripts/                 # Build & Infrastructure tools

The Patch Workflow

We follow the "Upstream First" philosophy used by Brave and Arch:

1. Dev: Make changes in a Chromium checkout.
2. Capture: Run scripts/create_patch.sh <category> <name> to generate a .patch.
3. Commit: Check in the patch file.
4. Build: scripts/bootstrap.sh applies patches in order to a clean Chromium tree.

---

🛠️ Building Ember

Prerequisites

• macOS: Xcode 15+, Python 3, Depot Tools
• Windows: VS 2022, Win SDK 10.0.22621+
• Linux: Ubuntu 22.04+, Clang 18+
• Hardware: 16GB RAM recommended for linking; ~100GB disk space.

Quick Start

1. Bootstrap Environment (Fetches Chromium & Dependencies):

   ./scripts/bootstrap.sh

2. Build Release Binary:

   ./scripts/build_all.sh --config release --platform mac

3. Run Tests:

   ./scripts/ci/validate_all.sh

---

🧠 AI Model Training

Ember's intelligence comes from a custom training pipeline that teaches the model to understand the DOM.

• Data Source: Synthetic generation via Playwright + Claude.
• Pipeline: scripts/training/pipeline.js handles crawling, EEP generation, replay verification, and dataset formatting.
• Fine-Tuning: LoRA/QLoRA on H100s via scripts/training/train.py.
• Quantization: Automated conversion to GGUF format optimized for Apple Silicon (Metal) and CUDA.

---

🔒 Security & Privacy

• DOM Sanitization: All HTML passed to the model is sanitized and token-limited to prevent prompt injection.
• Sandboxed Execution: AI actions run in an isolated context, separate from the main renderer.
• Manifest V3: Full compatibility with Chrome extensions, enforcing modern security practices.
• Update System: Custom update infrastructure separating Ember from Google's binary blobs.

---

📜 License

Copyright © 2026 Ember Browser Config. All rights reserved. CONFIDENTIAL & PROPRIETARY. Do not distribute.