Tencent BrowserSkill: the bsk CLI and a Chrome or Edge extension that let shell agents drive your logged-in browser

One bsk binary, a background daemon and an extension do the work together

Local automation needs three parts: an AI agent, the bsk CLI, and the browser extension. The CLI includes the background daemon, and a skill file teaches the agent which bsk commands to run. Any agent that can run shell commands uses the same CLI; named examples include Codex, Claude Code, Cursor, OpenClaw, CodeBuddy, WorkBuddy, Pi and Hermes Agent. DeepSeek Harness is the exception, with a dedicated plugin that exposes native browser tools and task previews. The CLI is written in Rust in a Cargo workspace with bsk-cli and bsk-protocol crates at version 0.4.1, requiring Rust 1.85, while the extension and plugins live in a pnpm workspace. The CLI and extension are released separately, as cli-v0.4.1 and ext-v0.4.1 on 2026-11-01.

Only Chrome and Edge on Chromium 125 or later are supported

The extension supports Chrome and Microsoft Edge based on Chromium 125 or later. Other Chromium browsers may work, but compatibility is not guaranteed, and Firefox and Safari are not on the list at all. The CLI and daemon run on macOS for Apple Silicon and Intel, on Linux for x64 and ARM64, and on Windows for x64 only, so Windows on ARM is outside the supported set. The extension installs from the Chrome Web Store or Edge Add-ons. For someone whose daily browser is Firefox, the consequence is that BrowserSkill cannot use the logins they already have, which removes the main reason to use it.

The agent acts with your real logins, inside a separate Agent Window

BrowserSkill works through the browser's existing login state, so the agent can read documents, search internal sites and complete web workflows as you. Its actions include reading page text and controls, clicking, typing, filling forms, dragging, selecting ranges, managing tabs and taking viewport or full-page screenshots; file upload and download work in local mode only. Work happens in its own Agent Window rather than in your current tabs. Existing tabs are used only when explicitly borrowed, and they are returned when the task ends. You can share control, pause the agent while you edit, and let it reconcile your changes before it continues. The trade-off is plain: anything your accounts can do, the agent can do. Name browser instances or bind a task to a specific profile to limit which sessions it sees, and enable the separate operation audit if you want a record of task activity, since that audit has to be switched on.

Setup can be handed to the agent, but the extension is still manual

The shortest route is to send the agent one instruction:

Set up browser-skill on this machine by following https://raw.githubusercontent.com/Tencent/BrowserSkill/main/AGENT_INSTALL.md

That guide covers installing the CLI, choosing the right skill or DSH plugin, connection checks and a first browser task. The agent cannot install the browser extension for you; that step stays with the user. Manual installation uses a shell script on macOS and Linux:

curl -fsSL https://raw.githubusercontent.com/Tencent/BrowserSkill/main/install.sh | sh
export PATH="${BSK_INSTALL_DIR:-$HOME/.local/bin}:$PATH"

and a PowerShell script on Windows:

irm https://raw.githubusercontent.com/Tencent/BrowserSkill/main/install.ps1 | iex

Both pipe a remote script straight into the shell, so read install.sh or install.ps1 first if your environment requires review. After installing, open the extension popup, enable the local connection and check its status once the CLI is running.

An agent that was already running will not find bsk until it restarts

The binary goes into ~/.local/bin by default, or into BSK_INSTALL_DIR when that is set. Verify it from the same terminal or agent environment that will call it:

bsk --version

One failure case is spelled out: an agent session that started before installation keeps its old PATH, so it cannot find bsk even though the terminal can. Either restart the agent so it picks up the new PATH, or configure the installed binary's absolute path in the agent.

Debugging evidence ties each action to requests, responses and console output

Beyond automation, BrowserSkill records evidence. It connects actions to network requests, response bodies, console messages and page changes, and inspects performance, slow APIs and suspected duplicate requests. A hypothesis can be tested with explicit HTTP rules or request replay. Debugging history stays local to the browser and can be reopened or exported as JSON. One authorized task tab can be recorded as a silent MP4 and previewed, saved or deleted from the extension. Loaded tables and lists export to JSON or CSV with column names, page and frame sources, merged-cell metadata and explicit coverage. Element waits take an explicit timeout, and the step command runs one chosen action and synchronizes, while view searches a saved observation without capturing the browser again.

Editorial conclusion

BrowserSkill fits developers who already work with a shell-capable agent and want it to act inside sites they are signed in to, debug a failing web flow with request and console evidence, or export tables, without setting up a separate automated browser. It is the wrong tool for Firefox or Safari users, and for anyone unwilling to let an agent act with their real sessions. Before installing, decide which browser profile the agent may use, turn on the operation audit if you need a record, and check that the agent process can see bsk on its PATH.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. README
4. Releases
5. Tencent/BrowserSkill on GitHub