Jonathon's AI Wiki

Claude in Chrome — Browser Automation Inside Your Logged-In Sessions

9 min read

Source: Claude Code docs (Chrome, CLI reference, best practices, Remote Control); Claude Help Center “Get started with Claude in Chrome” (updated 2026-08-26); Claude Code changelogs and release notes v2.1.154 to v2.1.282; the Notion Cowork recipe; four YouTube transcripts (“The expanding toolkit”, “8 Use Cases Every Marketer Should Know”, “30 Claude Features”, “7 NEW Use Cases”). Comparison rows: what’s-new W13/W14, browser-harness.com, Anthropic’s harness-design post and the Claude prompting best-practices doc.

Claude in Chrome is Anthropic’s browser extension that lets Claude “read, click, and navigate websites alongside you.” You can start its tasks from the Chrome side panel, from Claude Cowork or from Claude Code. Claude opens new tabs in your own browser and “shares your browser’s login state”, so it can work in any site you’re already signed into without an API connector. The Claude Code v2.1.198 release notes (2026-07-01) announced it as generally available, but the Help Center still lists the side panel as beta.

Key Takeaways

  • Three ways in: the Chrome side panel, the Claude Desktop connector, and Claude Code (--chrome, /chrome or VS Code).
  • Paid plans only: Pro, Max, Team and Enterprise. Claude Code needs a /login sign-in on a direct Anthropic plan. API-key, setup-token, Bedrock, Agent Platform and Foundry sessions can’t use it.
  • Your logins are its reach, so site permissions are the safety boundary. Claude pauses at login pages and CAPTCHAs and hands them to you.
  • Since v2.1.251, every browser action in Claude Code goes through Claude Code’s permission checks.
  • Best fits: AI-search tracking, reviews of competitor and SEO pages, data entry into tools with no API, and UI verification. Anthropic’s best-practices doc says “UI changes can be verified using the Claude in Chrome extension.”

Where You Can Use It

  • Chrome side panel. Click the Claude icon in the toolbar; Claude can read, click, type, navigate and fill forms on the current page. On Max and Team it runs as a Cowork session, so conversations are saved and skills and connectors work. This is rolling out to Pro; Enterprise needs an admin to enable Cowork in the cloud first. Recorded workflows exist only in the classic panel. Shortcuts can be scheduled from the clock icon.
  • Claude Desktop. Settings → Connectors → Claude in Chrome → Configure, then toggle it on. The connector is “disabled by default”, so you enable it in each conversation. Cowork is also getting a built-in browser. Existing Claude in Chrome users keep Chrome as their preferred browser, set in Settings > Cowork > Preferred browser.
  • Claude Code. Run claude --chrome, or choose /chrome → “Enabled by default”. VS Code has it whenever the extension is installed. Most sessions a Remote Control server creates start with Chrome off unless the server was started with --chrome (v2.1.273 or later).

Requirements and Known Limits

  • Extension version: Claude Code needs version 1.0.36 or later.
  • Browsers ^[ambiguous]: the sources disagree.
    • The Claude Code docs name Chrome and Edge, and say Claude Code “also detects the extension” in Brave, Arc, Vivaldi and Opera. It does not work under WSL.
    • The Help Center says it is “not supported on other Chromium-based web browsers or mobile devices.”
    • The “30 Features” video (fetched 2026-06-05) says it does not work in Brave, Arc or on mobile.
  • Context: enabling it by default loads the browser tools into every session, so the docs suggest --chrome only when needed.
  • Uploads: up to 10 MB per upload (v2.1.211 or later). Claude refuses files with multiple hard links, and files a rule denies it permission to Read.
  • Failure modes: a JavaScript alert or confirm dialog blocks commands until you dismiss it. If the extension goes idle in a long session, run /chrome → “Reconnect extension”.

Permissions and Safety

  • Anthropic’s own warning: the Help Center says it is “enhanced with our safety classifiers but is still risky.”
  • Site permissions are set in the extension (which sites Claude can browse, click and type on), and Claude Code inherits them. Team and Enterprise admins can turn the extension off or restrict sites with allowlists and blocklists. Claude Code follows the admin setting from v2.1.260.
  • Prompts: Claude Code’s dialog starts “Claude in Chrome wants to” and can allow all actions on that site for the session. Cowork asks “Allow Claude to browse the web on [site name]”, with the options “allow for this website / allow this time / allow always”.
  • Approval modes: the side panel starts in “Automatically approve”, where Claude reviews each action for safety and pauses when approval is needed. The “7 NEW Use Cases” video recommended “Ask before acting” over “Act without asking”.
  • Plan and auto mode: in plan mode, Claude asks before recording a GIF, opening a tab or running a shortcut. In auto mode, calls the classifier approves skip the extension’s per-site check (v2.1.275), unless your rules deny a site to Claude in Chrome.
  • Visibility: Claude works in a visible window and keeps its tabs in a tab group for the session. Cowork highlights active Claude tabs in orange.
  • Practitioner advice: the “30 Features” video says to keep it off banking and medical portals. The docs warn that GIF recordings capture account details on logged-in pages.

Release History (Claude Code)

Version (date)Change
v2.1.154 (2026-05-28)Choose among connected browsers via /chrome → “Select browser…“
v2.1.198 (2026-07-01)Generally available
v2.1.199 (2026-07-02)Plan mode now prompts for state-changing browser calls; stopped repeated reconnect pages
v2.1.211 (2026-07-15)Fixed file uploads from remote and CLI sessions
v2.1.221 (2026-08-04)Closes tabs it opened once they are no longer needed
v2.1.239 (2026-08-21)/clear closes the session’s tab group
v2.1.251 (2026-08-28)All browser actions go through Claude Code’s permission checks, even with telemetry off
v2.1.260 (2026-09-03)Follows the organization’s admin setting
v2.1.275 (2026-09-17)Auto mode skips the per-site check for classifier-approved calls
v2.1.282 (2026-09-24)allowClaudeInChromeWithManagedMcp managed setting

Uses for Marketers and Operators

  • AI-search visibility tracker (“8 Use Cases”). Cowork runs a list of target prompts in Perplexity, ChatGPT and Google AI Mode. It builds an Excel tracker for week-over-week comparison; the run can go in the background and took “fifteen to twenty minutes”.
  • Competitor sweep (“30 Features”). Ask it to “visit five competitor websites and return a structured table of their pricing, features, and positioning.”
  • SEO page review (Notion recipe). Claude reads Search Console data, opens “the top 5 performing pages using the web browser” and lists the fixes each page needs.
  • UX audit (“7 NEW Use Cases”, shown right after its Claude in Chrome setup). Cowork opened the presenter’s app in a browser, took the one-time passcode from their email, walked the new-member flow, and returned an audit in under 15 minutes.
  • Data entry and documents (Claude Code docs). Typing CSV rows into a CRM, drafting in a Google Doc, and extracting product listings to CSV.
  • Visual QA. In Anthropic’s demo, Claude Code reproduced a broken “new” button in Chrome, fixed the code and retested. The docs add form tests on localhost:3000 and checks against a Figma mock.

Claude in Chrome vs Playwright MCP vs Computer Use vs Browser Harness

ToolWhat it drivesSetupPick it when ^[inferred]
Claude in ChromeYour browser through Anthropic’s extension; shares your login stateExtension, then --chrome or the Cowork connectorThe work sits in a site you’re signed into and you want per-site approval
Playwright MCPA browser over MCP; Anthropic’s harness evaluator clicked through a running app “the way a user would”MCP server; listed as a verification tool in the prompting guideScripted QA of your own app, or sessions (API key, Bedrock, Foundry) where Chrome integration is off
Computer UseNative desktop apps, such as the iOS SimulatorDesktop Settings toggle (W13), or /mcp → computer-use in the CLI (W14, research preview)The target isn’t a web page
Browser HarnessYour real Chrome over one CDP websocket; MIT, Python, and the agent writes its own helpersuv tool install -e . plus Chrome remote debuggingYou want an editable open-source harness or hosted cloud browsers

Try It

  1. Install Claude (by Anthropic) from the Chrome Web Store. Sign in, pin it, and allow only the sites your first task needs.
  2. Run claude --chrome and ask Claude to search code.claude.com/docs for “hooks” (the docs’ test prompt). /chrome should show “Status: Enabled”.
  3. In Cowork, turn on the connector and run the competitor-sweep prompt against five real competitors.
  4. After your next UI change, ask Claude to open the local build, use the changed feature, and report any console errors.

Implementation

Tool/Service: the Claude in Chrome extension, used from Claude Code, Claude Desktop or the side panel. Setup: the first --chrome launch installs a native messaging host file. If the extension isn’t detected, restart Chrome. Cost: included in paid plans. The sources give no separate price. Integration notes: the tools are the claude-in-chrome MCP server (/mcp → View tools). Admins can block the server with deniedMcpServers. allowClaudeInChromeWithManagedMcp (v2.1.282) lets it run alongside an exclusive managed-mcp.json.

claude --chrome                  # enable for this session
claude --no-chrome               # disable for this session
claude remote-control --chrome   # Remote Control server sessions (v2.1.273+)

Open Questions

  • Brave and Arc support: the Claude Code docs and the Help Center disagree.
  • When does the side panel leave beta, given Claude Code’s GA on 2026-07-01?
  • Do browser tasks count against plan usage differently? Not stated.
  • How Playwright MCP handles your logins: none of the cited sources say.
  • Not fetched (two-page limit): the Help Center’s safety, permissions-guide and admin-controls pages.

Graph View

Backlinks