Purpose
Browy is a browser agent that runs on the user's existing AI subscription (GitHub Copilot). It installs as a Chromium extension and adds two interfaces: a sidebar chat and a DevTools panel CLI, both of which share session state, model selection, and chat history. Users can drive real, logged-in tabs through chat, letting the agent read pages and perform actions such as clicking and typing.
Main Capabilities
- Drive real tabs: read page accessibility tree snapshots, click and type by index, avoiding fragile CSS selectors or XPath.
- DevTools awareness: network requests, console logs, cookies, storage, and evaluate_js are all built-in tools, suitable for debugging an app you are building.
- Bring your own model: choose from the models offered by your Copilot plan (Claude, GPT, Gemini, etc.), switching via /model in the DevTools CLI or in settings.
- Single subscription: built on the GitHub Copilot SDK, using your existing Copilot plan (including Copilot Free); prompts count toward plan usage, the same as Copilot in your editor or terminal.
- Open source: Apache-2.0 licensed, with the agent loop in src/agent/loop.ts and the tool registry in src/agent/tools/browser.ts.
How to Use
- Install the extension from the Chrome Web Store (one click, covering Chrome, Edge, Brave). Pin the toolbar icon and click it; the sidebar opens and prompts you to install the backend.
- Install the native host: Windows provides an installer bundling Node and the Copilot CLI (about 140MB); macOS/Linux run the corresponding command. The sidebar automatically detects the backend within a few seconds.
- Sign in to GitHub Copilot: click sign in in the sidebar, a terminal window opens the device flow link, paste the code and authorize.
- Start chatting: send instructions to any tab and watch the tool calls and streaming replies.
Use Cases
- Scrape the YC startup directory to list company information in batches.
- Review GitHub PRs, summarize changes, and flag risks.
- Set up Gmail filters.
- Fill out signup forms across multiple tabs (multi-step, multi-tab, stopping before submission).
Limitations and Security
- Currently an early access v0.1.x preview; APIs, tools, and on-disk architecture may change between minor versions.
- Long multi-step automation on hostile single-page apps (such as LinkedIn, Notion, Discord, Figma) requires better waiting and retry strategies.
- Translations other than Simplified Chinese have not yet begun.
- The DevTools CLI is keyboard-only; mouse selection in the REPL is still imperfect.
- Tools that touch the host (Shell, filesystem, web_fetch, run_script) are off by default and must be explicitly enabled one by one in settings.
- No background activity: the agent acts only when the user sends a message, with no polling and no scheduled runs.
- Page content is sent to GitHub Copilot, the same path as running copilot in the terminal, and is subject to the Copilot subscription terms.
- Chat history is stored locally on disk in ~/.copilot/.
- Chrome shows the standard debugger banner while the agent is driving.stopping).
Limitations and Security
- Currently an early access v0.1.x preview; APIs, tools, and on-disk architecture may change between minor versions.
- Long multi-step automation on hostile single-page apps (such as LinkedIn, Notion, Discord, Figma) requires better waiting and retry strategies.
- Translations other than Simplified Chinese have not yet begun.
- The DevTools CLI is keyboard-only; mouse selection in the REPL is still imperfect.
- Tools that touch the host (Shell, filesystem, web_fetch, run_script) are off by default and must be explicitly enabled one by one in settings.
- No background activity: the agent acts only when the user sends a message, with no polling and no scheduled runs.
- Page content is sent to GitHub Copilot, the same path as running copilot in the terminal, and is subject to the Copilot subscription terms.
- Chat history is stored locally on disk in ~/.copilot/.
- Chrome shows the standard debugger banner while the agent is driving.
