BrowserSkill

BrowserSkill is a browser automation project open-sourced by Tencent, with the goal of enabling AI agents to work in a real browser where the user is already logged in, without interrupting the user's own browsing. It consists of two parts: the bsk CLI and a browser extension. The CLI contains a background daemon, and the extension is responsible for connecting to Chrome or Microsoft Edge.

Purpose BrowserSkill connects AI agents to Chrome or Microsoft Edge and uses the accounts the user is already logged into. Users can ask the agent to read pages, fill out forms, complete operations on websites, capture long screenshots, or troubleshoot failed requests. Tasks run in a separate, visible Agent Window; existing tabs can also be explicitly borrowed and returned after the task ends.

Main capabilities

• Browser tasks: read pages, fill out forms, operate websites, full-page long screenshots, investigate failed requests.
• Website debugging: start capture before reproducing an issue on a website authorized for debugging, then view the operation timeline, requests, console output, form field values, page changes, performance metrics, and API analysis; supports request filtering, JSON field inspection, in-task request modification, blocking, mocked responses, same-origin replay, and can save JSON records.
• Full-page screenshots: implemented through the extension's Quick Actions → Full-page screenshot for automatic scrolling, manual scrolling long screenshots, or visible-area screenshots; this extension feature can be used without an agent or daemon connection.
• Browser profile selection: open the extension in the target profile and copy the profile instructions, then send them with the task; you can also specify a unique name for the browser and select it explicitly.
• Remote operation: the browser and login state remain on the local computer, while the agent, CLI, and daemon can run on a server, paired through authenticated WSS; the browser actively initiates the connection, so no inbound port is needed locally.

How to use Local automation requires three parts: AI agent + bsk CLI + browser extension. The CLI contains the background daemon, and the skill is used to teach the agent how to use it. Install the CLI (Windows PowerShell): irm https://browser-skill.org/install.ps1 | iex The default binary is installed under ~/.local/bin, and this needs to be checked in the terminal or agent environment that will use it. If a running agent cannot find bsk, restart it to pick up the new PATH, or configure the absolute path to the installed binary. Install the extension: install it from the Chrome Web Store or Edge Add-ons, open the popup, enable the local connection, and check the status after starting the CLI. Install the skill: use bsk install-skill, select the harness with the spacebar, then press Enter; for non-interactive installation, specify it explicitly. Use bsk install-skill --list to view supported targets and paths. Existing installations are skipped unless explicitly replaced with --force. For other harnesses, you can copy the entire crates/bsk-cli/skill/ directory into their skills directory as browser-skill/, including references/. DSH users should install the plugin, which already includes the skill. Connection check: resolve failed check items and confirm that the extension shows Connected. Start a new agent session and confirm that it can discover browser-skill; a successful doctor check alone does not mean the skill has been discovered. First task: after connecting, you can have the agent open its own window, read the page, return a summary, and stop the task. External links or manually created tabs that fall into the task window are moved to the most recently used normal window; if there is no normal window, a user window is created. Routing happens after tab creation, so a tab may briefly appear in the task window. Tabs opened by the page, manual assistance, or recording are preserved. Harnesses that support skill slash commands can also use /browser-skill. Session management: start a session and keep the returned session_id, then use that value in place of <id> in subsequent commands. When connecting to multiple browsers, you can select one at session startup with --browser <instance-id-or-name>. By default it is created in the background; --no-focus is still supported, and --focus explicitly brings the new window to the foreground. Requesting manual help and interactive recording will still bring the relevant window to the foreground. Use bsk --help or bsk <command> --help to view command options. After the task is complete (including failed tasks) should always stop the session, and borrowed tabs are returned to their original window.

DeepSeek Harness Plugin The DSH plugin adds native bsk_* tools, browser task previews, and screenshot results to the DeepSeek Harness Web UI, uses the same bsk CLI and extension, and comes with the BrowserSkill skill. When upgrading to 0.3.3, you need to update the CLI and extension first, then update the plugin; the six browser_* tools have been renamed to bsk_*, no aliases under the old names are provided, so you need to update custom tool calls and policies, restart the DSH profile, and re-invoke /browser-skill in existing conversations.

Use cases

• Let the agent read, fill in, and operate on websites where you are already logged in.
• Reproduce issues and collect evidence on websites authorized for debugging.
• Need full-page long screenshots or visible-area screenshots.
• Remote deployments where the agent runs on a server while the browser and login state remain local.
• Work with shell-capable agents such as Cursor, Claude Code, Codex, OpenClaw, CodeBuddy, WorkBuddy, Pi, Hermes Agent, and others.

Limitations and cautions

• Agent Window shares the login state of the selected profile; it is not a separate account or a security sandbox; the agent can act with the permissions of the logged-in website, so tasks and agents must be chosen carefully.
• The extension has two independent Automation settings, both enabled by default; these browser settings apply to existing and new sessions, and the legacy --unattended, tab-borrowing --no-confirm, and BSK_REQUEST_HELP=off inputs cannot override them. Disabling human help does not mean the requested steps have been completed. Closing Agent Window can stop its tasks.
• Chrome may still activate the window when a page opens a native popup; Chrome's public extension APIs cannot guarantee zero momentary activation, nor can they guarantee restoring focus to other applications.
• Full-page screenshots follow the document scroll area, and there are limitations with nested scroll panels and virtualized lists.
• Request replay sends new requests using the page's current session, which may change server data. Capture and redaction have limitations, and retained evidence may still contain sensitive information; known secrets are filtered from debugging evidence, but redaction cannot guarantee removal of all sensitive data; video frames are notAutomatic redaction.
• Stopping a task or disconnecting will not delete retained history or videos; exported copies and copies already received by agents or gateways must be managed separately.
• BrowserSkill does not operate a mandatory cloud service and does not collect product telemetry; automation results are sent to the selected daemon or gateway and to the agent using it, and that agent or service may process or retain data according to its own policies. The extension does not independently call AI providers.
• bsk supports Windows 10 and Windows Server 2016 or later.
• Updates: Complete active browser tasks first, then update the CLI. The default local installation installs the new executable and checks that it runs while the daemon continues serving, then restarts the daemon on the same port with the new executable; if the restarted daemon is not ready, it restores the old executable and restarts from it. A daemon started with --foreground keeps the old version running and must be restarted in its terminal or supervisor. The background daemon checks for new versions every 30 minutes and installs them in the same way when there are no active agent sessions. Daemon-side automatic updates can be disabled with BSK_AUTO_UPDATE=off while retaining manual bsk update. The extension updates through the browser store; the DSH plugin must be updated separately and its profile restarted.
• Source build: The repository is a Rust + pnpm workspace and requires Rust stable, Node.js 22, and the pnpm version declared in package.json. The CLI builds under target/release/; load apps/extension/dist/chrome-mv3/ as an unpacked extension. Extension development uses pnpm ext:dev. Relevant checks can be run with cargo test --workspace --locked, pnpm ext:test, and pnpm lint.
• The extension interface supports English, Simplified and Traditional Chinese, Korean, Japanese, French, Italian, Spanish, German, and Brazilian Portuguese.

The project is licensed under MIT.