Browy in Action: Zero-Cost AI Browser Agents via GitHub Copilot

Introduction: The Hidden Cost of AI Browser Agents

AI browser agents face an underappreciated engineering problem: token consumption.

Every time an agent interacts with a page, it sends the DOM structure (typically 3000–5000 tokens) to an LLM for decision-making, then performs the next action. A moderately complex task — filling a form, navigating, extracting data — easily burns through tens of thousands of tokens. With paid API pricing, a single complex task can cost over $1 in inference.

Dario Amodei, CEO of Anthropic, noted in 2025 that agentic models break the traditional compute economy, with platforms potentially consuming $8 in inference compute for every $1 of subscription revenue.

Browy sidesteps this problem in an elegant way: it routes all model calls through the GitHub Copilot SDK. If you already have a GitHub Copilot subscription (personal $10/mo, enterprise $19/mo), every browser automation task costs zero in marginal inference.

This article covers Browy's architecture, installation, dual-interface operation, and practical workflows.

Architecture: Extension + Native Host + Copilot

Browy's architecture has three layers:

┌─────────────────────────────────────┐
│        Chrome Extension Layer       │
│  ┌─────────────┐  ┌──────────────┐ │
│  │  Side Panel │  │ DevTools CLI │ │
│  └──────┬──────┘  └──────┬───────┘ │
│         │                │          │
│         └──────┬─────────┘          │
│                ▼                    │
│         chrome.debugger API         │
│         (Chrome DevTools Protocol)  │
└──────────────────┬──────────────────┘
                   │
                   ▼
┌─────────────────────────────────────┐
│     Native Messaging Host           │
│   (Node.js daemon, runs locally)    │
│                                     │
│   ┌──────────────────────────┐      │
│   │  GitHub Copilot SDK Wrapper      │
│   └──────────┬───────────────┘      │
│              │                      │
│   ┌──────────▼───────────────┐      │
│   │  41 Built-in Tool Registry       │
│   └──────────────────────────┘      │
└──────────────────┬──────────────────┘
                   │
                   ▼
┌─────────────────────────────────────┐
│     GitHub Copilot API              │
│  (Claude, GPT-4o, Gemini, etc.)     │
└─────────────────────────────────────┘

Extension layer: A standard Chrome extension that attaches to the active tab via chrome.debugger, intercepting CDP traffic.

Native host layer: A local Node.js process communicating with the extension through Chrome's native messaging protocol. The host wraps the GitHub Copilot SDK, routing all model requests through the Copilot API.

Model layer: The Copilot subscription provides fixed-rate access to frontier models including Claude, GPT-4o, and Gemini. No separate API key required.

Key Design Decisions

Installation

Three steps: extension, native host, Copilot login.

Step 1: Install the Extension

Add "Browy" from the Chrome Web Store (also works with Edge and Brave).

Verify: Pin the toolbar icon and click it. The side panel shows a "host disconnected" banner — expected until Step 2.

Step 2: Install the Native Host

macOS / Linux:

# Fetch and run the install script from Browy's project
curl -fsSL https://example.16yun.cn/browy-install.sh | bash

Windows (PowerShell):

# Fetch and run the install script from Browy's project
irm https://example.16yun.cn/browy-install.ps1 | iex

This installs:

• The native host binary at ~/.browy/app/ (macOS/Linux) or %LOCALAPPDATA%\Browy\app\ (Windows)
• The native messaging manifest so Chrome can launch the host process
• A sandboxed scratch directory at ~/.browy/data/files/

Verify: Reopen the side panel. The host banner should be gone.

Step 3: Sign in to GitHub Copilot

Click "Sign in to GitHub Copilot" in the side panel. A terminal opens with a device-flow link. Paste the code in your browser, authorize, and the terminal closes.

Verify: Send "what is the headline of this page." You should see a snapshot tool call followed by the model reply.

Two Interfaces: Side Panel vs DevTools CLI

Side Panel Daily Operations

Open any page and type natural language instructions:

Extract all links from this page and format them as a table

Find the search box, search for "best AI agents 2026," then extract the top 5 results

DevTools CLI Power Features

Slash commands for keyboard-driven workflows:

# Switch model in DevTools CLI
/model claude-sonnet-4
 
# Debug with JavaScript
/js document.querySelectorAll('.price').map(e => e.textContent)

Core Toolset

Browy includes 41 tools covering browser operations.

Browser Tools

Host OS Tools (Disabled by Default)

These touch the host operating system and are disabled by default for security:

To enable, go to Browy Settings and toggle each tool individually.

Practical Examples

Example 1: Scrape YC Startup Directory

List all YC Fall 2026 companies with their location and one-line pitch

Browy flow:

1. Load YC startup page
2. snapshot to get page structure
3. Identify each company card in the list
4. Extract: company name, location, description
5. Format as table output

Example 2: Review a GitHub PR

Summarise the changes in this PR and flag anything risky

Browy:

1. Reads the PR diff view
2. Analyzes changed files
3. Extracts key code changes
4. Applies Copilot's code review capability

Example 3: Set Up a Gmail Filter

Auto-archive everything from no-reply@*.atlassian.net

Browy navigates Gmail settings → Filters → Create new filter and fills the form.

Security Model: Default-Off

Browy uses a default-off security strategy:

• All OS-level tools (file read/write, terminal) are disabled by default
• Users enable them individually in Settings
• This prevents prompt injection from malicious websites

User visits site → Malicious site injects instruction →
┌─────────────────────────────────────┐
│  Browy security barrier:            │
│  - Host tools default off           │
│  - Each tool must be manually enabled
│  - File I/O sandboxed               │
└─────────────────────────────────────┘
         ↓ Attack fails

Using Proxies

Browy runs in the user's real browser, inheriting all existing sessions and cookies. For scraping scenarios requiring IP rotation, configure at the browser level.

Method 1: Chrome startup with proxy flag

google-chrome --proxy-server=http://user:pass@proxy.16yun.cn:8888

Method 2: System environment variable

export HTTP_PROXY=http://user:pass@proxy.16yun.cn:8888
export HTTPS_PROXY=http://user:pass@proxy.16yun.cn:8888
google-chrome

Browy vs Alternatives

Limitations

• No headless: Requires a graphical desktop browser; cannot run on servers
• No concurrency: Single-user, single-tab tool; not for large-scale extraction
• Copilot dependency: Copilot's model availability and rate limits affect Browy directly
• China access: api.githubcopilot.com is reachable but slow from mainland China

Summary

Browy's innovation is finding a specific gap — the existing GitHub Copilot subscription that many developers already pay for — and exploiting it to perfection.

By routing model calls through a fixed-rate subscription, Browy drives the marginal inference cost of browser automation to zero. Combined with its dual-interface design, default-off security, and 41 built-in tools, it's a practical and safe daily automation tool for developers.

The next article covers Steel Browser — a fundamentally different architecture: a Docker-based cloud browser API designed for large-scale concurrent AI agents.