Skip to content

README_EN.md

Latest commit

 

History

History
201 lines (137 loc) · 9.72 KB

README_EN.md

File metadata and controls

201 lines (137 loc) · 9.72 KB
Raw

中文 | English

Sabrina

Sabrina

Unified memory · Shared skills · 90% of your daily computer use, covered

Stars License Platform Electron Download

OpenClaw through IM alone is incomplete.
Sabrina is the browser presence it's been missing.

OpenClaw handles IM. Sabrina handles the browser.
Same skills. Same memory. One AI system.

Sabrina — Browser + AI Sidebar

What It Does

OpenClaw users spend a huge chunk of their day in the browser — researching, reading docs, checking competitors, organizing information. In all of those moments, OpenClaw's IM channel can't help.

Sabrina closes that gap: every page in your browser becomes a direct entry point into the full capabilities you've already built in OpenClaw. Skills don't need reconfiguring. Memory doesn't break. Your model policies keep working.

Open any webpage — the sidebar already knows what you're looking at. The page is the input. OpenClaw is the engine.

Common use cases:

  • 🤖 Agent mode — Describe a task in plain language; Sabrina operates the actual browser tab for you: clicking, filling forms, navigating, scrolling — up to 20 steps autonomously. High-risk actions (form submissions, deletions) pause for your confirmation; everything else runs unattended. The sidebar shows a live step-by-step journal with screenshots, AI reasoning, and risk ratings for each action
  • ⚡ Skills straight to the browser — Trigger any OpenClaw skill directly on any webpage. Reading competitor docs? File an issue. Reviewing a contract? Generate a summary. Browsing a codebase? Run your review skill — page content becomes the input automatically
  • ✨ Coding GenTab — Select multiple reference tabs; Sabrina runs a three-pass pipeline to generate a self-contained interactive HTML page: design planning → full HTML/CSS/JS generation → automatic QA with error auto-fix. The result is a real usable artifact, not a table, not a chat log
  • 🤝 Handoff — Hand a task off to OpenClaw to run asynchronously in the background, with the current page as context. Keep browsing — results come back when it's done
  • 🧵 Memory follows the page — Conversation history auto-archives by page and site, reusing your existing OpenClaw memory conventions. Next time you open the same page, context is still there
  • 📎 Multiple tabs as context — Reference several open tabs at once and feed the full information density of your browser directly to OpenClaw — not one paste at a time

How It Compares

Sabrina Tabbit Sider / Monica / Extensions BrowserOS / Dia / AI Browsers ChatGPT / Claude Web
Context source Auto-reads current page + selection + multi-tab refs @mention tabs, groups, files, screenshots Manual select or copy Partial auto, often screenshot-based Fully manual paste
Browser automation ✅ Agent mode — natural language drives real browser actions, with risk gating ✅ Background Agent (cross-tab) ❌ ⚠️ Limited ❌
Multi-tab collaboration ✅ Cross-tab refs + Coding GenTab ✅ @group refs + background agent ⚠️ Single page only ⚠️ Limited ❌
AI capability source Reuses your existing OpenClaw stack Built-in multi-model (GPT / Gemini / Claude etc.) Self-contained closed system Self-contained closed system Platform-locked
Thread continuity ✅ Auto-associated by page/site, persists across sessions ❌ No session persistence ❌ Isolated per conversation ⚠️ Partial ❌ Isolated
Connection Local / SSH / HTTP Endpoint / Relay pairing code Cloud Extension Built-in Browser
Open source ✅ MIT ❌ Closed freeware ❌ ❌ ❌

Sabrina doesn't reinvent AI — it lets your existing OpenClaw work natively in the browser.

Quick Start

Download & Install (Recommended)

→ Releases page — download the latest .dmg

⚠️ Current builds are unsigned. On first open: right-click → Open, or run: xattr -cr /Applications/Sabrina.app

Build from Source

git clone https://github.com/jiaqi015/openclaw-ai-browser.git
cd openclaw-ai-browser
npm install
npm run dev

Prerequisites: macOS + Node.js 18+ + OpenClaw installed locally or remotely

Connect OpenClaw

  1. Open Sabrina → OpenClaw settings
  2. Choose connection mode:
    • Local — OpenClaw running on this machine, connect directly
    • SSH remote — enter SSH address, executes remotely
    • HTTP Endpoint — enter endpoint URL + access token for a remote service
    • Relay pairing — generate a 6-character code, connect remote OpenClaw
  3. Run a quick health check — you're ready

See Connecting to OpenClaw for details.

Features

🤖 Agent mode — Natural language drives real browser actions. Click, fill forms, navigate, scroll — up to 20 steps; high-risk actions pause for confirmation; live sidebar journal shows screenshots, reasoning, and risk rating per step.

✨ Coding GenTab — Multi-tab content → interactive HTML page. Three-pass pipeline: design planning → code generation → auto QA; JS runtime errors auto-fixed.

🔍 Zero-friction page context — Open the sidebar and Sabrina already knows what you're looking at.

🗂️ Multi-tab references — Reference multiple open tabs simultaneously; feed the full information density of your browser session to OpenClaw.

⚡ Skills in full context — OpenClaw's skill ecosystem works directly in the browser, with page content as natural input.

🤝 Handoff — Move a task to OpenClaw for async background execution with current page context attached.

🔄 Real-time model switching — Switch models mid-task without leaving the browser.

🧵 Thread memory — Conversation history auto-archives by page and site, reusing OpenClaw memory conventions across sessions.

🔌 Four connection modes — Local CLI, SSH remote, HTTP Endpoint, Relay pairing code — works in any network environment.

💡 Why Sabrina

Sabrina is not "yet another AI browser."

It is OpenClaw's native workspace for the browser — bringing OpenClaw's existing agents, skills, memory, model policies, and runtime sessions into the richest, highest-frequency work surface on your computer.

Most AI tools ask you to leave the page first, then rebuild context in a chat box. Sabrina inverts this:

  • No copying links and selections to "feed" the AI
  • No re-describing what you're already looking at
  • No interrupting browser work before engaging AI

The page you're looking at is the most important input.

Sabrina's advantage isn't rebuilding an AI platform — it's reusing OpenClaw's established capability layer: agent, auth, model policy, skill ecosystem, session conventions. Different surface. Same capabilities.

🏗️ Architecture

Three layers: Browser UI → Main Process → OpenClaw (via pluggable driver)

flowchart LR
  subgraph Sabrina["Sabrina (Electron)"]
    UI["Browser UI\nTabs · Nav · AI Sidebar"]
    Main["Main Process\nTabs · Threads · State"]
    Context["Page Context\nSnapshot · Selection · Multi-tab refs"]
    Agent["Browser Agent\nAutonomous ops · Risk gating"]
  end

  subgraph Driver["Transport Layer"]
    D1["local-cli"]
    D2["ssh-cli"]
    D3["endpoint"]
    D4["relay-paired"]
  end

  subgraph OpenClaw["OpenClaw"]
    OCA["Agent / Brain"]
    Eco["Models · Skills · Memory"]
  end

  UI -->|IPC| Main
  Main --> Context
  Main --> Agent
  Context -->|context package| Driver
  Agent -->|hands mode| Driver
  Driver --> OCA
  OCA --> Eco
  OCA -.->|response| Main
  Main -->|render| UI
Loading

Docs

Doc Contents
Connecting OpenClaw Setup steps for all four connection modes
Browser Agent Design Agent mode: risk gating, Brain-Hands architecture
Turn Engine Design Turn lifecycle, execution planning, receipt normalization
GenTab PRD Full GenTab spec including Coding GenTab pipeline
Engineering System Architecture boundaries and conventions
Design Baseline UI tone, component constraints, extension rules
System State Current system overview, what's real, what's next

Contributing

PRs and Issues welcome. Read Engineering System first to understand architecture boundaries, then run npm run acceptance to make sure nothing regressed.

If Sabrina is useful to you, a ⭐ is the best way to support it.

License

MIT