# BrowserSkill

> BrowserSkill connects AI agents like Cursor, Claude Code, and Codex to your real, logged-in browser via a CLI and browser extension, enabling browser automation without interrupting your work.

BrowserSkill is an open-source project from Tencent that lets AI agents automate your real, already-logged-in browser without requiring separate test accounts or interrupting your normal browsing. It works through a local CLI (`bsk`) paired with a browser extension, bridging any shell-capable AI agent to Chrome or Microsoft Edge. The latest release is CLI v0.3.0, published in September 2026.

## What It Is

BrowserSkill is a local browser automation bridge designed for AI agent workflows. Instead of spinning up a headless browser or requiring agents to log in separately, it reuses your existing browser session. The agent issues commands through the `bsk` CLI, which routes them via a local daemon to the BrowserSkill browser extension, which then executes tasks in a dedicated Agent Window — keeping your normal browser windows untouched unless a tab borrow is explicitly requested and approved.

## How the Architecture Works

The system has three local runtime components working in sequence:

- **`bsk` CLI/daemon** — receives shell commands from the agent, manages sessions, and communicates with the extension over a local WebSocket on `127.0.0.1`
- **Browser extension** — installed in Chrome or Edge, automates the Agent Window, and handles tab borrowing with user confirmation
- **Skill file** — a `SKILL.md` instruction set installed into the agent harness (Cursor, Claude Code, Codex, OpenClaw, CodeBuddy, WorkBuddy, Pi, Hermes Agent, DeepSeek Harness, and others) that teaches the agent how to invoke `bsk`

The agent never communicates with the browser directly. DeepSeek Harness users get a dedicated npm plugin (`@wxg-prc-cpg/browser-skill-dsh-plugin`) that injects native `browser_*` tools and a live session view in the Web UI.

## Agent Harness Compatibility

BrowserSkill is explicitly model- and framework-agnostic. Any agent that can call a shell command can use it. The README lists first-class support for Cursor, Claude Code, Codex, OpenClaw, CodeBuddy, WorkBuddy, Pi, and Hermes Agent, with a dedicated plugin path for DeepSeek Harness. Skills can be installed with a single `bsk install-skill` command, and custom skill files are supported for teams with their own agent instructions.

## Human-in-the-Loop Controls

A key design feature is built-in human oversight. When an agent encounters a CAPTCHA, login prompt, confirmation dialog, or other step that requires human action, it can call `request-help` to pause and hand control back to the user, then resume afterward. Two independent automation settings in the extension popup control this behavior:

- **Confirm before borrowing tabs** — requires user approval before the agent touches any of your open tabs
- **Allow requests for human help** — enables or disables the `request-help` mechanism

Both settings are saved per browser profile and apply immediately to all active and new sessions.

## Platform Support and Setup

BrowserSkill runs on macOS (Apple Silicon and Intel), Linux (x64 and ARM64), and Windows x64. The browser extension is available on the Chrome Web Store and Microsoft Edge Add-ons store, with support for other Chromium-based browsers via unpacked extension install. Firefox support is listed as planned.

Installation follows a three-step path: install the `bsk` CLI via a shell script or PowerShell command, install the browser extension from the store, then run `bsk install-skill` to register the skill with the target agent harness. A `bsk doctor` command verifies the full connection and reports any configuration issues.

## Update: CLI v0.3.0

The latest release (CLI v0.3.0, September 2026) introduced a notable behavioral change: the `--unattended`, `tab borrow --no-confirm`, and `BSK_REQUEST_HELP=off` flags no longer bypass confirmation or disable help requests. These controls moved to the extension's Automation settings UI, making the browser profile's saved preferences authoritative. The release also added protocol 1.3, which enforces `request-help` routing through the daemon rather than allowing local CLI exits. The repository is a Cargo + pnpm workspace with the codebase split across Rust crates (CLI/daemon, protocol types) and TypeScript packages (extension, UI, i18n, DeepSeek plugin).

## Features
- Reuse real browser login state without separate test accounts
- Dedicated Agent Window keeps normal browser windows untouched
- Tab borrowing with explicit user confirmation
- Built-in human-in-the-loop via request-help mechanism
- Works with any shell-capable AI agent (Cursor, Claude Code, Codex, OpenClaw, etc.)
- bsk CLI and local daemon for session management
- Browser extension for Chrome and Microsoft Edge
- Full-page screenshot support
- DeepSeek Harness native plugin with browser_* tools
- Remote browser connections via built-in authentication service
- Sandboxed agent setup support
- bsk doctor health check command
- Skill auto-update management with local edit preservation
- Localization: English, Simplified Chinese, Korean

## Integrations
Cursor, Claude Code, Codex, OpenClaw, CodeBuddy, WorkBuddy, Pi, Hermes Agent, DeepSeek Harness, Chrome, Microsoft Edge, npm (dsh plugin)

## Platforms
CLI, BROWSER_EXTENSION, WINDOWS, MACOS, LINUX, API

## Pricing
Open Source

## Version
cli-v0.3.0

## Links
- Website: https://github.com/Tencent/BrowserSkill
- Documentation: https://github.com/Tencent/BrowserSkill/tree/main/docs
- Repository: https://github.com/Tencent/BrowserSkill
- EveryDev.ai: https://www.everydev.ai/tools/browserskill