Claude in Chrome

What it is

A Chrome extension that opens Claude in a side panel alongside whatever you’re browsing. Claude reads the page, and — when you ask — clicks, types, switches tabs, and takes actions on your behalf, gated by a permission system that asks before anything risky.

It’s available on all paid plans (Pro, Max, Team, and Enterprise). Anthropic describes Claude in Chrome as generally available in Cowork and Claude Code, and in beta in the Chrome side panel. Chrome only — no other Chromium browsers, no mobile.

Where to find it

Visit the Chrome Web Store listing.
Click Add to Chrome.
Sign in with your Claude account.
Pin the extension: puzzle-piece icon → the thumbtack next to Claude.
Grant site permissions when prompted.

The Claude icon then sits in your Chrome toolbar; click it to open the side panel.

Its settings live at Settings → Claude in Chrome, on claude.ai and in the Mac app alike. The pane is headed Claude in Chrome settings and holds three things:

an Enable Claude in Chrome switch — “Use Claude in your browser with the Claude in Chrome extension. This setting only affects the extension”;
a Site permissions section, which notes that “These permissions apply to Claude in Chrome and the built-in browser in the Claude desktop app”. Its one control is a Default for all sites dropdown — “Choose whether Claude works on all sites by default”. Set to Allow all sites, the pane reads “Claude works everywhere except sites you block below”;
a Blocked sites table — “Claude cannot be used on these sites” — with an Add websites button.

Two side panels

Which panel you get depends on your plan:

Max and Team — the side panel runs as a Cowork session, so it uses the same interface as Cowork on the web and desktop and your conversations are saved across devices. This is rolling out to Pro accounts.
Enterprise — you get the Cowork panel once an owner has enabled Cowork in the cloud; until then the classic panel.
Anywhere the Cowork panel is active, Switch back to classic in the menu returns you to the old one.

How to use it

Pick a permission mode

A drop-down on the chat input offers three modes:

Manually approve (Manual) — Claude pauses before each action and you choose Allow or Deny. This was previously called “Ask before acting”.
Automatically approve (Auto) — Claude keeps working and screens its own actions for safety, blocking unsafe ones and pausing when it needs you. This is the default in the Cowork side panel, and it uses more of your usage allowance than the other modes.
Skip all approvals (Skip) — no pauses and no automatic safety checks. Previously called “Act without asking”. Use it only when you trust every site, connector, and file involved.

When Claude asks about a sensitive site you can choose Allow this action for that one action, or Always allow actions on this site for an ongoing grant. Approved sites are listed under Your approved sites in the extension settings.

Core flows

Ask Claude about the page you’re on. Select text, share a screenshot region, or upload an image to point at something specific.
Drag tabs into Claude’s tab group to let it read across several at once.
Save a good prompt as a shortcut and type / in chat to reuse it. Shortcuts can be scheduled to run on a recurring basis.
Record a workflow — perform the steps yourself and Claude learns the pattern so it can repeat it. This is a classic-panel feature.
Claude has enhanced navigation for Slack, Google Calendar, Gmail, Google Docs, and GitHub, and can read console logs and network requests when you’re debugging.
From Claude Desktop: enable the Claude in Chrome connector (Settings → Customize → Connectors → Claude in Chrome) to hand browser tasks to the extension without switching windows.
From Claude Code: build in the terminal, then test and verify in Chrome — the extension gives Claude console logs, network requests, and DOM state.

Claude in Chrome runs on all the public models available to your plan.

Good to know

Prompt injection is the headline risk. Malicious instructions hidden in a page can try to hijack Claude. Anthropic’s testing reports that Claude Opus 4.8 is substantially more robust than earlier models and that its current configuration reduces attack success rates to less than 0.08% in internal testing — low, but not zero. Start on trusted sites, watch for sudden topic shifts, and stop the task if something feels off.
Some things Claude won’t do at any permission level: make purchases or financial transactions, create accounts, handle sensitive payment or identity data, download from untrusted sources, perform permanent deletions, give financial advice or place trades, or modify system files.
Blocked and gated sites. Claude can’t access adult-content sites or known pirated-content sites. Financial sites require your permission before Claude goes near them.
Even “always allow” has limits. After granting a site ongoing permission, Claude still asks before downloading files, entering sensitive information, or granting authorisations.
Use a separate browser profile without access to your sensitive accounts, and don’t leave the panel open over confidential pages — screenshots of the tab are sent to Claude while it’s open.
Admin controls (Team and Enterprise): Organization settings → Claude in Chrome, for an Owner or Primary Owner. There’s an org-wide enable switch, an allowlist and a blocklist layered over the default blocked categories, and a separate toggle — off by default — for 1Password credential filling on macOS. Team plans are on by default; Anthropic’s admin guide says Enterprise plans are off by default and become enabled on 10 September 2026 unless an owner has already disabled them. Enterprise owners can also disable the Desktop-to-Chrome connector with isLocalDevMcpEnabled.
You control the extension from the Mac app too. The extension itself only runs in Chrome, but the Claude desktop app carries the same Settings → Claude in Chrome pane — the Enable Claude in Chrome switch, the Default for all sites dropdown, and the Blocked sites table are all there, so you don’t have to go to claude.ai to change them. The pane adds that these site permissions “apply to Claude in Chrome and the built-in browser in the Claude desktop app”, so the same allow-and-block list governs both.
Usage counts against your plan — the same pool as Claude and Claude Code. Browser work uses more tokens per turn than plain chat.
Troubleshooting: Claude can’t see the page → refresh and check you’ve granted permission for that site. Actions misbehaving → update Chrome and disable extensions that also manipulate the page. No Cowork panel → check your plan, wait for the Pro rollout, or ask your admin to enable Cowork.

What’s changed recently

2026-09-13 — The Beta badge has gone from the settings pane. claude.ai/settings/browser-extension is headed Claude in Chrome settings with no badge on the heading, and the Claude in Chrome row in the settings nav of the Mac app carries none either — we recorded that badge in April and May, and it is no longer there. The pane’s own controls have been rewritten to match the dump: Site permissions is a section heading, not a control — the control under it is a Default for all sites dropdown whose value on the swept account was Allow all sites — and the Blocked sites table takes entries through Add websites. Our previous “Allow extension” wording is gone; “Default for all sites” turns out to have been right all along.
2026-09-13 — Two material changes. The side panel now runs as a Cowork session on Max and Team plans, is rolling out to Pro, and reaches Enterprise once an admin enables Cowork in the cloud; Switch back to classic returns you to the old panel, and workflow recording stays a classic-panel feature. The permission modes have been renamed and there are now three — Manually approve (Manual) (was “Ask before acting”), Automatically approve (Auto) (new, and the default in the Cowork panel), and Skip all approvals (Skip) (was “Act without asking”). Also updated: the prompt-injection figure is now Claude Opus 4.8 at under 0.08% attack success; banking and brokerage sites are permission-gated rather than blocked (only adult and pirated-content sites are blocked outright); the model list is “all public models available to your plan”, replacing the old per-plan list we had; and 1Password credential filling on macOS is a new, off-by-default admin toggle. Claude in Chrome no longer appears in Anthropic’s beta and research preview table, though its own help articles still describe the Chrome side panel as beta.