BrowserSkill

BrowserSkill connects a shell-capable AI agent to the browser you are already signed in to, and keeps the agent's work out of your way. Browser tasks run in a separate, visible Agent Window; a tab you already have open is touched only when the agent borrows it explicitly and returns it when the task ends.

It has two local pieces: the bsk command-line tool with its daemon, and a browser extension. The agent never reaches the browser directly — it shells out to bsk, the daemon routes the request over a loopback WebSocket to the extension, and the extension performs the action in the Agent Window. Any agent that can run a shell command can therefore use it, with no dependency on a particular model or framework.

A skill file teaches the harness how to call bsk. Running bsk install-skill selects and installs it for Cursor, Claude Code, Codex, OpenClaw, CodeBuddy, WorkBuddy, Pi, or Hermes Agent, and copying skill/SKILL.md into a harness's skills directory covers the rest. DeepSeek Harness instead uses a published npm plugin that injects native browser_* tools and adds a live web overlay of each Agent Window, without shelling out to bsk.

When a task reaches a step only a person can complete — a captcha, a login, a confirmation dialog — the agent can hand control back to you and continue afterwards. The CLI and extension run on macOS, Linux, and Windows, with Chrome and Microsoft Edge supported and other Chromium browsers expected to work where they accept unpacked extensions.

Features

• Logged-in session reuse: agents work against sites you are already authenticated to, without separate test accounts
• Separate Agent Window: automation happens in its own visible window so your normal browsing continues uninterrupted
• Explicit tab borrowing: an existing tab is used only when the agent asks for it, and is returned when the task finishes
• Shell-based integration: any agent that can call a shell command works through the bsk CLI, with no framework lock-in
• One-step skill install: bsk install-skill installs the skill into a selected harness, with manual copying for others
• DeepSeek Harness plugin: an npm plugin injects native browser_* tools and a live observation overlay instead of shelling out
• Human-in-the-loop: captchas, logins, and confirmation dialogs hand control back to you mid-task
• Cross-platform: macOS on Apple Silicon and Intel, Linux on x64 and ARM64, and Windows x64
• Browser support: Chrome and Microsoft Edge, with Chromium browsers that accept unpacked extensions expected to work
• Local transport: the CLI, daemon, and extension communicate over local IPC and a loopback WebSocket