Tabhand

[圖片]

Tabhand

Let your AI agents use your own Chrome, signed in to your accounts,
in background tabs that never steal your focus.

tabhand.dev · Docs · Chrome Web Store

[npm version] [Chrome Web Store] [MIT license] [macOS and Windows: Chrome, Brave, Edge, and Dia on macOS]

https://github.com/user-attachments/assets/d5fe5c5a-c012-471f-bf3b-10fa754c382d

_{Video not playing? Watch it on tabhand.dev}

npx tabhand setup

Works with Claude Code, Codex, Cursor, OpenCode, Gemini CLI, Copilot and more: 19 agents in all, plus anything else that speaks MCP.

Then add the extension from the Chrome Web Store. Setup opens the page for you. It works in Brave, Edge and Dia too.

Why Tabhand

• Your logins, not a blank browser. Agents work in the Chrome you already use, so email, dashboards and admin panels just work. You never hand over a password.
• You keep working. Every agent gets tabs of its own in the background, grouped under its name and muted. Chrome never jumps to the front.
• Your tabs stay yours. An agent can only reach the tabs it opened, never yours and never another agent's.
• You can watch. Each agent has its own cursor that glides, clicks and types, and its tabs wear its icon. The Tabhand button in Chrome's toolbar shows who's working and pauses every agent at once.
• No servers. Tabhand runs entirely on your computer, with no analytics. Pages only go to the agent you're running. Read the privacy policy.

[Claude's cursor working on a Wikipedia page in a background tab]

How it works

agent ──stdio──▶ tabhand mcp ──Unix socket──▶ native host ──native messaging──▶ extension ──chrome.debugger──▶ agent's tabs

• Background by default. Each agent session opens its own tabs with active: false, inside a tab group named after the agent. Focus emulation keeps those tabs rendering, so clicks, typing and screenshots work without bringing anything to the front. Agent tabs are muted, and the agent moves on once a page is usable instead of waiting for every video to load.
• Agents stay in their lane. The extension only lets a session reach tabs that session opened, plus any tabs and popups those pages open. Your tabs, and other agents' tabs, are invisible to it.
• Gets along with password managers. Chrome won't let Tabhand control a page while it shows another extension's frame, such as iCloud Passwords' autofill menu. Tabhand takes those frames out of agent tabs and carries on with the step it was on.
• No network port. The bridge talks over a Unix socket in a directory only you can read, with a fresh token each run. Web pages and other machines can't reach it.
• Nothing sensitive logged, no telemetry. The bridge logs connect and disconnect events only: no URLs, page content or typed text.
• You can watch. While an agent works in a tab, its cursor sits on the page, even across page loads. It glides to whatever the agent clicks, ripples on each click, shows typing dots while it types, waves when you switch to the tab, and fades out once the agent is done. It only animates while you're looking, and it's hidden from the agent's own screenshots.
• Agent tabs stand out. While an agent works in a tab, the tab's icon becomes the agent's arrow. The site's own icon comes back when the agent is done.
• Each agent wears its own colors, or yours. Claude, Codex, OpenCode, Cursor, Gemini and Copilot each get their own cursor and tab group color; anything else gets a violet cursor with its own name. To give every agent your team's name, color and icon instead, click the Tabhand icon in Chrome's toolbar, then the gear, and choose Custom. The tag's icon can be the name's first letter, one of ten marks, or none, in white or black to suit the color. To set it for every agent on the computer, run tabhand cursor --name Acme --color '#FF5A1F' --icon initial.
• Stop button. Click the Tabhand icon in Chrome's toolbar and press Pause every agent, or click Cancel on Chrome's "started debugging" bar. Either pauses every agent at once.
• Several Chrome profiles. If Tabhand is in more than one profile (say, work and personal), agents see each one and can ask for it: "use my Work profile". Click the Tabhand icon in a profile, then its name at the top of the panel, to give it a name agents know it by, or fill in that profile's Google email. Without a choice, agents use the profile that connected first. Pause Tabhand in one profile and new work goes to a profile where it's on. An agent that asked for a profile stays in it: if you pause that one, the agent hears so rather than signing in somewhere else.

[Each agent gets its own cursor: Claude, Codex, OpenCode, Cursor, Gemini, Copilot]

[Tab badges for each harness]

Setup (macOS and Windows)

npx tabhand setup

That's all:

• It installs Tabhand in ~/.tabhand and connects it to Chrome, and to Brave, Edge and Dia (macOS) if you have them.
• It adds Tabhand to every agent it finds, for all your projects. Each new session gets the browser tools, and the server's instructions tell the agent when to reach for them.
• It asks before changing anything. Where an agent has its own mcp add command, setup uses it. Otherwise it edits the agent's settings file in place, keeping comments, and backs up the original to ~/.tabhand/backups.
• It opens Tabhand in the Chrome Web Store, so you can add the extension (once), and waits for Chrome to connect.

Agents it knows:

• Claude Code, Codex, OpenCode, Cursor, Gemini CLI, Qwen Code, VS Code (Copilot), Copilot CLI, Kiro, Amp, Factory Droid, Warp, Junie, Antigravity
• Devin Desktop (Windsurf), Claude Desktop, Cline, Zed

To add it to only some of them, run npx tabhand setup --harness claude,codex.

Using Brave, Edge or Dia? Add the extension there from the same Chrome Web Store page. Edge asks you to allow extensions from other stores first. Each browser shows up to agents as one more profile: click the Tabhand icon in it and give it a name, such as "Brave", so agents can ask for it. Arc isn't supported: tab groups and screenshots never finish there, so Tabhand stays out of Arc even if you add the extension. Arc shuts down on December 22, 2026; Dia works.

tabhand status checks the install, the Chrome connection and each agent. tabhand remove undoes it all. Run npx tabhand@latest setup again to update.

Any other MCP client can run ~/.tabhand/bin/tabhand mcp itself (on Windows, %USERPROFILE%\.tabhand\bin\tabhand.cmd mcp).

On Windows

• You need Node.js 20 or later, from nodejs.org or winget install OpenJS.NodeJS.LTS.
• In PowerShell, if npx says running scripts is disabled on this system, run npx.cmd tabhand setup instead, or use Command Prompt.
• Setup registers Tabhand with Chrome and Edge (Edge comes with Windows), and with Brave if you have it, under HKEY_CURRENT_USER in the registry. Nothing needs an administrator.
• Agents reach the bridge over a named pipe, not a network port, and each has to show a token kept in your own AppData, new each time Chrome starts the bridge.
• Tested on Windows 11 with Chrome, and with Claude Code as the agent.

Working on Tabhand

npm installnode bin/tabhand.mjs setup    # builds dist/tabhand.mjs and installs it from this checkoutnode bin/tabhand.mjs reload   # after changing the extension's code

From a checkout, setup has you load the extension unpacked (Developer mode, then "Load unpacked") instead of adding it from the Chrome Web Store.

Setup copies a single bundled file (dist/tabhand.mjs, built by npm run build), because macOS doesn't let Chrome's child processes read Documents, Desktop or Downloads. Agents and Chrome both start it through ~/.tabhand/bin/tabhand. That launcher remembers which Node.js ran setup, since agent apps don't get your shell's PATH. If that Node.js is removed later, it finds another one. If Chrome never connects, the bridge's log at ~/Library/Application Support/tabhand/host.log says why.

On Windows, Chrome starts the bridge through ~\.tabhand\bin\tabhand-host.cmd, which the registry points it to, and agents run the Node.js that ran setup with ~\.tabhand\app\tabhand.mjs directly, since they start programs without a shell. The log is in %LOCALAPPDATA%\tabhand\host.log, and anything the bridge printed while failing to start is in host-errors.log next to it.

Windows-only code lives in src/windows.mjs and extension/windows.js. The shared files call into them only behind a Windows check, and keep their macOS code as it is.

Tools

Tests

• npm test runs all of these except the smoke test.
• node test/setup.mjs runs setup, status and remove in a throwaway home folder holding existing agent configs. It checks that Tabhand gets added, that comments and other settings survive, that a second run changes nothing, and that the launcher serves the MCP server.
• node test/setup-windows.mjs does the same on Windows, with the registry entries under a key of the test's own, and checks that agents' own mcp add commands get called through cmd.exe.
• On Windows the tests that start a browser need TABHAND_TEST_BROWSER set to Chrome for Testing's chrome.exe.
• node test/package.mjs packs the npm package, installs the tarball, and runs setup from it.
• node test/activator.mjs checks, on macOS, that the helper the bridge starts to hand the front back to your app goes away with the bridge, even when the bridge is killed, and that a new bridge stops helpers an older Tabhand left running and nothing else.
• node test/arc.mjs checks, on macOS, that a bridge Arc starts stays connected but out of agents' way.
• node test/profiles.mjs runs two bridges against stand-in extensions, "Work" and "Personal". It checks that agents see both by name and can choose one, that a chosen profile stays chosen, and that pausing one profile sends new work to the other.
• node test/profile-name.mjs names a throwaway Chrome for Testing profile in the toolbar panel and checks that agents see the name, and that the panel pauses and resumes agents.
• npm run test:e2e launches a throwaway Chrome for Testing profile, runs two agents in parallel, and checks they can't reach each other's tabs and that the browser never comes to the front.
• node test/new-tabs.mjs checks that links, window.open() and popups from an agent's tab never take over your tab, and that the agent gets the new tab.
• node test/drag-hover.mjs checks, in a background tab, that hover-only buttons (shown by CSS :hover or by a mouseenter handler) can be reached, and that lists reorder by dragging, with the browser's own HTML5 drag and drop and with mouse events alone, including lists that want a press delay and ones that have to be scrolled while an item is held.
• node test/upload.mjs checks that files reach a page byte for byte, through a file field or a button's file picker, including files over Chrome's 1 MB message cap.
• node test/foreign-frames.mjs checks that an agent gets through pages where a password manager's autofill menu pops up, using a stand-in extension.
• npm run test:smoke runs a short check against the Chrome you have open.
• node test/live.mjs --profile Dia --app Dia runs every tool in a real browser profile you have Tabhand in (Dia, Brave, Edge or a Chrome profile), through your install: two agents at once on a local test page, checking that their tabs stay apart and that the browser never comes to the front.
• node test/cursor-presence.mjs checks the cursor and tab badge show before the first action, survive page loads, and leave when the agent is done.
• node test/load-speed.mjs [--real] compares how long the agent waits for pages with a plain tab.
• node test/cursor-gallery.mjs renders every cursor theme side by side, and node test/badge-preview.mjs every tab badge.

Roadmap

• Per-site approval: agents ask before using a site for the first time
• Confirmation before sending, buying or deleting
• Logins filled from your password manager without the AI seeing the password
• Downloads and iframes
• Linux

License

MIT