"Don't Just Browse. Command."
The AI-native browser that acts as your agentic companion.
Highlights · Modes · Getting Started · Architecture · Contributing
Currently built on Zen Browser 1.19.6 and Firefox 149.0
Nevoflux is not just another browser. It's your personal AI companion that lives in your browser — it has its own identity, its own soul, and it learns from every interaction with you.
Built on Zen Browser (Firefox/Gecko), Nevoflux bridges the gap between you, AI, and the internet. It browses with you, thinks with you, and acts for you — from simple conversations to autonomous web navigation and full computer control.
Under the hood: a Rust-native agent daemon with 16+ LLM providers, 80+ browser automation APIs, cross-platform computer control, MCP integration, an encrypted cross-session memory system, and integration with external coding agents and personal AI assistants.
- 🧬 Identity & Soul — Your companion has its own personality, values, and behavioral patterns that evolve over time
- 🧠 Memory & Learning — Remembers across sessions, learns your preferences, adapts to how you work — with encrypted storage
- 🎛️ Progressive Autonomy — Three modes (Chat, Browser, Agent) — you decide how much control to hand over
- ⚡ Build: Micro Apps (Canvas) — Generate fully functional mini-apps on the fly — not just rendered artifacts, but living apps with full agent capabilities. Exportable to 10+ formats (HTML, PNG, PDF, DOCX, PPTX, XLSX, SVG, ZIP). Every other AI just shows you text — Nevoflux creates tools you can actually use.
- 🐙 Manage: OpenClaw Integration — Your personal AI assistant that manages your life beyond coding — schedules, tasks, knowledge, and workflows. Nevoflux is the unified interface: Canvas Micro Apps help you build tools, OpenClaw helps you manage everything else.
- 🧩 WASM Skills — Extensible skill system powered by WebAssembly — sandboxed, pluggable, and progressively loaded
- 🖥️ Browser + Computer Control — 80+ browser APIs and cross-platform desktop control (screenshot, mouse, keyboard)
- 🤖 Multi-LLM Freedom — 16+ providers: Anthropic, OpenAI, Qwen, DeepSeek, Gemini, Ollama, and more — your choice
- 🔧 Coding Agent Delegation — Seamlessly delegate development tasks to Claude Code, Gemini CLI, or Kimi Agent — they work as sub-agents within your browser
Nevoflux uses progressive capability unlocking — each mode builds on the previous one, giving you control over how much autonomy your companion has.
Chat — Talk to your companion. Ask questions, search the web, let it read what's on your screen.
Browser — Let it take the wheel. It clicks, types, navigates, and fills forms for you.
Agent — Full autonomy. File operations, shell commands, computer control, sub-agents — your companion handles complex multi-step tasks end-to-end.
| Mode | Capabilities |
|---|---|
| Chat | LLM reasoning, memory, web search, read-only browser access |
| Browser | Chat + click, type, navigate, fill forms, interact with pages |
| Agent | Browser + file ops, shell commands, computer control, MCP, sub-agents |
| Provider | Type | Notes |
|---|---|---|
| Anthropic Claude | API | Default (claude-sonnet-4) |
| OpenAI | API | GPT-4o series |
| Google Gemini | API | Up to 1M context window |
| Qwen (通义千问) | API | DashScope integration |
| DeepSeek | API | |
| Ollama | Local | Run local models |
| Groq | API | |
| Mistral | API | |
| XAi | API | |
| Cohere | API | |
| Perplexity | API | |
| Together | API | |
| OpenRouter | API |
| Agent | Type | Notes |
|---|---|---|
| Claude Code | Coding | Subprocess integration |
| Gemini CLI | Coding | Subprocess integration |
| Kimi Agent | Coding | Subprocess integration |
| OpenClaw | Personal | Personal AI assistant integration |
Configure your preferred provider in ~/.config/nevoflux/config.toml:
[llm]
active_provider = "anthropic"
default_model = "claude-sonnet-4-20250514"Nevoflux consists of two components: the browser (this repository) and the Rust native agent (nevoflux-agent). Both are needed for full functionality.
# Prerequisites: Node.js 22+, npm, Python 3
git clone https://github.com/dorisgyl/nevoflux.git
cd nevoflux
npm install
npm run download
npm run bootstrap
npm run import
npm run build
npm run start# Prerequisites: Rust 1.75+
git clone https://github.com/dorisgyl/nevoflux-agent.git
cd nevoflux-agent
cargo build --releaseThen register the native messaging host:
# From the nevoflux (browser) repo
./scripts/setup-native-host.sh┌──────────────────────────────────────────────────────────────────────────┐
│ Entry Layer │
│ │
│ Browser Extension Claude Code (MCP) CLI │
│ (Native Messaging) (stdio JSON-RPC) (--daemon/--mcp) │
│ │ │ │ │
└─────────┼────────────────────────┼─────────────────────────┼──────────────┘
│ │ │
└────────────────────────┼─────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────────────────────┐
│ Daemon Layer (Rust) │
│ │
│ ┌──────────────┐ ┌───────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │ Agent Engine │ │ Context │ │ Session │ │ Permission │ │
│ │ (agentic loop)│ │ Builder │ │ Manager │ │ Enforcer │ │
│ └──────┬────────┘ └───────────────┘ └─────────────┘ └─────────────┘ │
│ │ │
│ ┌──────▼──────────────────────────────────────────────────────────────┐ │
│ │ Backend Services │ │
│ │ │ │
│ │ LLM Providers │ Browser Tools │ Computer │ WASM │ MCP │ │
│ │ (16+ services, │ (80+ APIs) │ Control │ Skills│ Server + │ │
│ │ local, agents) │ │ │ │ Client │ │
│ └──────────────────────────────────────────────────────────────────────┘ │
│ │
│ ┌──────────────────────────────────────────────────────────────────────┐ │
│ │ Memory & Learning (encrypted, cross-session) │ │
│ │ Knowledge Graph │ Conversation History │ User Preferences │ Skills │ │
│ └──────────────────────────────────────────────────────────────────────┘ │
│ │
│ ┌──────────────────────────────────────────────────────────────────────┐ │
│ │ SQLite Storage (WAL mode, encrypted) │ │
│ └──────────────────────────────────────────────────────────────────────┘ │
└──────────────────────────────────────────────────────────────────────────┘
Build ▼ Manage ▼
┌─────────────────────┐ ┌──────────────────────────┐
│ Canvas Micro Apps │ │ OpenClaw │
│ Create tools, │ │ Personal AI assistant │
│ export to 10+ │ │ — schedules, tasks, │
│ formats │ │ knowledge, workflows │
└─────────────────────┘ └──────────────────────────┘
Code ▼
┌────────────┐ ┌─────────────┐ ┌─────────────┐
│ Claude Code │ │ Gemini CLI │ │ Kimi Agent │
│ (coding) │ │ (coding) │ │ (coding) │
└────────────┘ └─────────────┘ └─────────────┘
Coding Agents
nevoflux (browser repo)
├── src/nevoflux/
│ ├── extensions/nevoflux-agent/ # WebExtension (sidebar UI, background.js)
│ │ ├── dioxus-ui/chat-sidebar/ # Dioxus/WASM sidebar application
│ │ ├── background/ # Extension background script
│ │ └── content/ # Content scripts for browser tools
│ ├── engine-overlays/ # Browser chrome modifications
│ │ ├── browser/actors/ # NevofluxChild/Parent actors
│ │ ├── browser/components/ # nevoflux:// pages (canvas, settings)
│ │ └── browser/modules/ # Bridge router, content store
│ ├── patches/ # Zen Browser patches
│ └── overlays/ # New files for src/zen/
The native agent is organized as a Rust workspace with 10 crates:
nevoflux-agent/
├── nevoflux-daemon # Core daemon (agent engine, session, memory, permissions, WASM runtime)
├── nevoflux-bridge # Proxy bridges (Native Messaging, MCP)
├── nevoflux-protocol # Message types and serialization (MessagePack + JSON)
├── nevoflux-llm # LLM provider abstraction (via rig-core)
├── nevoflux-storage # SQLite persistence (sessions, messages, memory, permissions)
├── nevoflux-mcp # MCP server + client integration
├── nevoflux-computer # Cross-platform computer control (X11, macOS, Windows)
├── nevoflux-skills # WASM skill loading and management
├── nevoflux-builtin-wasm # Built-in WASM agent module
└── nevoflux-testing # Test infrastructure (mocks, builders, helpers)
Nevoflux's Canvas system lets you generate interactive artifacts — HTML apps, React components, SVG graphics, Mermaid diagrams, multi-file projects, and more — directly in the chat.
Canvas artifacts are living applications with access to the NevofluxSDK, enabling them to call browser tools, interact with the agent, and access storage.
| Format | Availability |
|---|---|
| Source | Always |
| HTML | Always |
| PNG | Always |
| Always | |
| DOCX | Always |
| SVG | SVG / Mermaid artifacts |
| Markdown | HTML artifacts |
| PPTX | Slides artifacts |
| XLSX | Artifacts containing tables |
| ZIP | Project (multi-file) artifacts |
Nevoflux plays a dual role in the MCP ecosystem:
- As MCP Server — Exposes 23+ tools (browser, computer, agent) to external clients like Claude Code
- As MCP Client — Connects to external MCP servers (filesystem, GitHub, and more)
# ~/.config/nevoflux/mcp-servers.toml
[servers.filesystem]
command = "npx"
args = ["-y", "@modelcontextprotocol/server-filesystem", "/home/user"]
[servers.github]
command = "npx"
args = ["-y", "@modelcontextprotocol/server-github"]
env = { GITHUB_TOKEN = "..." }Nevoflux uses a default-deny permission model. Every sensitive action requires explicit approval.
- Human-in-the-loop — Security approval, quality control, decision branching, and error intervention
- Sensitive path blocklist —
.ssh/,.gnupg/,.aws/credentials, and other critical paths are always blocked - Permission scopes — Grant access once, for the session, or permanently
- Encrypted storage — Memory and knowledge data encrypted with AES-256-GCM
- SDK sandboxing — Canvas artifacts run in sandboxed iframes; the NevofluxSDK is only injected into
nevoflux://pages, never external websites
# ~/.config/nevoflux/config.toml
[daemon]
port_range_start = 19500
idle_timeout_secs = 1800 # 30 min idle shutdown
[llm]
active_provider = "anthropic"
default_model = "claude-sonnet-4-20250514"
max_tokens = 4096
temperature = 0.7
[session]
max_sessions = 500
inactive_cleanup_days = 90
[storage]
max_size_mb = 1024
wal_mode = trueAPI keys are resolved in order: environment variables → system keychain → config file.
Q: How do I open the AI sidebar?
Press
Ctrl+Shift+A(Windows/Linux) orCmd+Shift+A(macOS) to toggle the AI sidebar open and closed.
Q: How do I configure LLM providers and agents?
Type
nevoflux://settingsin the address bar and navigate to AI Models. You'll see two sections: LLM Providers (cloud APIs and local models) and Agents (coding agents and personal assistants). Click any card to configure its API key and settings.
Q: How do I connect Claude Code as a coding agent?
First install the Claude Code ACP adapter globally:
npm install -g @agentclientprotocol/claude-agent-acpThen go to
nevoflux://settings→ AI Models → click the Claude Code card and configure it.
Nevoflux is open source and welcomes contributions.
- Discord — Join our community
- Bug Reports — GitHub Issues
- Feature Requests — GitHub Discussions
- Code Contributions — See CONTRIBUTING.md
Nevoflux uses a patch-based system for customizing Zen Browser. The engine/ directory is regenerated from upstream + patches — never edit it directly.
# Development cycle
npm run import # Apply patches/overlays
# ... make changes in src/zen/ ...
npm run build:ui # Quick UI rebuild
npm run start # Test in browser
# Before committing
./scripts/export-nevoflux-patches.sh # Export changes as patches
./scripts/revert-zen-changes.sh # Revert src/zen/
git add src/nevoflux/patches/
git commit -m "patch(feature): description"| Command | Description |
|---|---|
npm run import |
Apply patches and overlays |
npm run build |
Full browser build |
npm run build:ui |
UI-only rebuild (faster) |
npm run start |
Launch the browser |
npm run reload-ext |
Reload extension + clear caches |
npm run test |
Run tests |
npm run lint |
Run ESLint/Prettier |
- Zen Browser — The productivity-focused browser that serves as our foundation
- Firefox — The open-source Gecko engine
- OpenClaw — Personal AI assistant framework
- Wasmtime — WebAssembly runtime
- Model Context Protocol — AI-tool integration protocol
Nevoflux is licensed under the Mozilla Public License 2.0.
Nevoflux — Don't Just Browse. Command.
This project is developed and maintained by YULIN GAN.
- GitHub: https://github.com/dorisgyl
- Contact: doris_gyl@hotmail.com