Claude Now Lives Inside Google Workspace: What It Does, What It Refuses, and What It Costs You

On 6 October, Anthropic shipped a public beta that puts Claude inside Google Docs, Sheets and Slides as a sidebar. It reads the file you have open, edits it in place, and records every change under your name in version history. It also works in the other direction: paste a Google file link into Claude and the file opens beside the chat.

The announcement matters for a reason that has nothing to do with model quality. For years, the practical argument for using Gemini inside Google Workspace was simply that it was already there, sitting closest to the files. That argument has now been answered, and the answer runs on a subscription many teams already pay for, delivered inside Google Workspace. What follows is the part the launch posts leave out: exactly what the extension can touch, what it refuses outright, and where the Google Workspace beta still gets in your way.

What Claude does inside the file

The mechanic is straightforward. One Marketplace install adds the Google Workspace sidebar to all three applications, not three separate extensions. When you send a message, Claude reads the relevant parts of the file you are looking at, plus anything you have selected, and makes its changes directly in the document, sheet or deck.

Three details make it more useful than a chat window bolted onto a page.

You can point at a specific part of the file. Select text in Docs, cells in Sheets, or a slide in Slides, and the selection defines the scope of the request.

It answers questions without editing anything. Asking what changed between two tabs or which slides mention pricing does not modify the file, which makes it usable as a reading tool rather than only a writing tool.

You control how much it asks first. The mode selector at the bottom of the sidebar governs whether Claude applies changes directly or checks with you. In Google Docs specifically, you can have it propose edits as cards that you accept or reject one at a time, which is close to the suggestion workflow reviewers already use.

You also choose the model from the sidebar, drawn from whatever your Claude plan offers.

Three ways Claude can reach your files, and which one you want

Anthropic now ships three overlapping ways for Claude to touch documents, and confusing them leads to wrong expectations.

The Google Workspace extension is the new one. It puts Claude inside a single Docs, Sheets or Slides file in Google Workspace and edits that file in place. Narrow, immediate, and the right tool when you are already working on the document.

The Google Workspace connectors run the other way. Turn them on and Claude at claude.ai can search and reference your Drive, Gmail and Calendar while you chat. That is the one to reach for when a question spans documents rather than living inside one of them.

Claude in Chrome is different again, working across web pages in your browser, which is useful for research and for pages that are not files at all.

The practical difference is scope. The extension sees one file. The connectors reach your Drive. Claude in Chrome sees what is on your screen. Pick the narrowest one that does the job, because each step up in scope is a step up in what an attacker could reach if a prompt injection landed somewhere it should not.

What it refuses to do, which is the more interesting list

Anthropic has drawn explicit lines, and reading them tells you what the company is worried about.

Three actions always stop and ask, regardless of mode, and are labelled as a security review: inserting an image from a web address, adding a link, and removing anyone's access. There is no always allow option for those three, which is deliberate, because each is a way a document could be turned into a delivery mechanism for something malicious.

Four things Claude will not do at all, even if asked directly. It cannot share the file. It cannot change ownership. It cannot create menus or scheduled automations. And it cannot load arbitrary web pages into your file.

That last set matters because it defines the blast radius. An assistant that can restyle a deck inside Google Workspace is convenient. An assistant that can share the deck, email a link, or run on a schedule is a security incident waiting for a mistake. Anthropic has shipped the first and refused the second, at least in beta.

The install reality for most companies

Here is where a lot of teams will hit a wall that has nothing to do with Claude.

Many organisations block Google Workspace Marketplace apps by default, and the first sign of it is a greyed out install button or a message reading that the application is not allowed by your administrator. That message comes from Google, not from Anthropic, and only a Workspace admin can clear it, either by allowlisting the app or by installing it centrally.

The good news for Google Workspace administrators is the scope. A super admin can install Claude for everyone, or for specific groups and organisational units, which means a pilot is a small configuration change rather than a company wide rollout. Removal is equally contained: uninstall it in the admin console and it disappears for all users the next time they reload a file.

One practical note from the documentation that saves a support ticket. Extensions only appear in files opened after the installation, so if Claude is missing from the Extensions menu, reload the file before assuming the install failed.

The limitations worth knowing before you rely on it

The Google Workspace beta carries a list of limits, and several are deliberate rather than unfinished, because lifting them would widen what the extension can reach.

• Six minute ceiling on any single operation. Google limits an extension action to six minutes, so a long job stops partway. The workaround is to ask Claude to continue, or to split the request, for example one tab or ten slides at a time.
• Connected Sheets are out of reach. A sheet backed by BigQuery through Connected Sheets cannot be read. The documented path is to extract the rows into a regular tab first.
• Chat history lives in your browser, per file. It will not follow you to a different browser, a private window, after clearing site data, or into a copy of the file made with Make a copy. Treat a conversation as disposable and put anything important into the document.
• Google formats only. Word, Excel and PowerPoint files are not supported, and the file has to be converted first.
• Browser matters. Chrome, Edge and Safari work. Firefox with strict tracking protection can block the sidebar from talking to the document.
• One Google account at a time. If your browser is signed in to more than one account, reading and editing may fail until the file is opened in a session with a single account.
• Claude account sign-in only. If your team reached Claude through an API key, Amazon Bedrock, Google Vertex AI or a gateway, this extension does not support that route.

Those limits are the reason the honest summary of this release is that it is excellent for editing work and poor for batch work. It is a collaborator, not a pipeline.

Security, privacy, and the one risk that is specific to documents

Three things in the documentation deserve to be read carefully by anyone who approves tools for a team.

The permissions the Google Workspace extension requests are narrow. The extension asks Google for two permissions, the minimum needed to display a sidebar and edit the file it is opened in. It does not request access to Drive, Gmail, Calendar or contacts, and it cannot make network requests from inside your file. By default, Claude sees only the file you have open. If you want it to reach other Drive files, that requires turning on the separate Google Drive connector, which is off by default.

That narrow permission set is worth contrasting with how most Google Workspace add-ons behave. Plenty of them ask for broad Drive access because it is easier to build that way. Anthropic has taken the harder path here, and the reason is not generosity: an extension that can only touch the open file is an extension whose failure modes stay small.

Your content is handled under your existing Claude terms and is not used for training. Administrators control who has the extension, and the broader enterprise controls apply as documented.

Prompt injection is the risk that comes with the format. A document can contain instructions written by somebody else, sometimes hidden in white text or a tiny font, and those instructions are exactly the kind of thing that would ask an assistant to share a file, follow a link or contact someone. Anthropic's stated defences are that document content is treated as data rather than commands, that hidden text is labelled when found, and that the assistant is designed not to follow a document's instruction to share, fetch or message. The practical advice is unchanged and applies to every AI assistant working inside files: review proposed actions before approving them, and be more careful with files that arrived from outside your organisation. We looked at how prompt injection has been used against coding assistants in our write-up on the Codex and Copilot breaches, and the pattern there is the same one to watch for here.

What this changes if you work in Google Workspace all day

Strip the announcement language away and three practical shifts are worth your attention.

The Gemini default is no longer the only convenient option. Any team that chose Gemini inside Google Workspace because it was already there now has a real alternative sitting in the same sidebar, running on their existing Claude subscription. For teams already paying for both, this is a straightforward productivity question rather than a platform decision.

Editing in place removes the worst part of the old workflow. The copy, paste, edit, copy back loop was where both quality and time went to die, and it is the specific problem this extension solves. If your team spends its day moving content between an AI chat and a document, the gain is immediate rather than theoretical.

Version history becomes the review surface. Because Claude's edits are recorded under your name, they appear in version history like any edit you make yourself, and the chat is private to your browser. That combination is better than it sounds: collaborators see the changes, not the prompts, and undo works the way it always has. The organisational risk is that reviewers will assume every edit under your name was made deliberately by you, when some of them came from an assistant that was told to restructure a section.

For anyone whose work is closer to content production than document editing, the same logic applies to other workflows, and we covered the assembly-line version of it in our guide to content repurposing. If your work is closer to code, how developers wire Claude and Codex into their daily flow is the more relevant piece, and the model itself keeps moving, as our notes on the Claude Opus release show.

The other shift worth naming is organisational rather than technical. When the assistant lives inside Google Workspace, the work it does is visible in the file rather than in a separate chat history nobody else can audit. For teams that have been uneasy about staff pasting client material into a browser tab, an extension that stays inside the document, attributes edits to the author and keeps the conversation on the local machine is a smaller governance problem than it first appears. That does not make it safe by default. It makes it reviewable, which is the property most workplace AI tools still lack, and it is the reason I would expect the Google Workspace extension to spread faster inside companies than an API integration ever does.

Six things to do before you switch a team on

1. Check the admin policy first. Confirm whether your Google Workspace organisation blocks Marketplace apps, and whether you are allowlisting the extension or installing it centrally.
2. Pilot with one team, not the company. The admin console supports group and organisational unit installs, which makes a contained test easy.
3. Agree the approval mode before anyone starts. Decide whether edits land directly or arrive as suggestion cards in Docs, and write the answer down.
4. Test on a real document, not a demo. Take the untidiest Google Workspace spreadsheet your team owns and hand it a genuine task, including one that will run past the six minute limit.
5. Set a rule for external files. Documents that arrived from outside the organisation are where prompt injection lives, and the review habit matters more there than anywhere else.
6. Check your connector settings. If the Google Drive connector is on for anyone, Claude can reach beyond the open file, which changes the data story you told your security team.

Where the claims stop

Four limits on the above.

• This article is built from Anthropic's own documentation and announcement. I read the documentation rather than running the extension on a paid plan, so the workflow details are as documented, not as tested.
• The Google Workspace extension is in beta, and beta documentation changes without notice.
• The security posture described is the vendor's own design intent. Prompt injection defences are a moving target, and no assistant working inside documents can be described as solved.
• The six minute operation cap is Google's platform limit, not something Anthropic can fix, which means it is the kind of constraint that quietly shapes how teams use the product.

FAQ

Is Claude for Google Workspace free?

It is in beta on all paid Claude plans, and it requires a Google Workspace account. The extension itself does not appear to carry a separate price, so the cost is your existing subscription plus an administrator decision to allow the app.

Can Claude see my other Google Drive files?

Not by default. The extension is restricted to the single file it is opened in. Reaching other files requires the separate Google Drive connector, which is off by default and worth checking before you assume anything about data access.

Is this the same as Claude in Chrome?

No, and the distinction matters. Claude in Chrome works across web pages in your browser. The Google Workspace connectors let Claude search your Drive, Gmail and Calendar from claude.ai. This extension is narrower and more useful: it puts Claude inside one specific Docs, Sheets or Slides file and lets it edit that file.

Should I use this or Gemini in Google Workspace?

Both can now sit in the same Google Workspace sidebar, so the question becomes which one your team already pays for and which answers your tasks better. The honest answer is that this release removes the convenience argument for Gemini rather than settling the capability question, and only a test on your own documents will settle that.

The short version

Claude for Google Workspace puts an AI sidebar inside Docs, Sheets and Slides, reads the file you have open, edits it in place and records the changes under your name. It refuses to share files, change ownership, run on a schedule or load arbitrary web pages, and three actions always require approval, which tells you the design was built around limiting the damage a document can do. The beta's real constraints are a six minute ceiling per operation, Google formats only, browser chat history and an admin gate that stops many teams at the install button. For anyone who already pays for Claude and lives in Google Workspace, it is the most practical thing Anthropic shipped this month.

Sources

• Anthropic, use Claude in Google Docs, Sheets and Slides: https://support.claude.com/en/articles/16951679-use-claude-in-google-docs-sheets-and-slides
• Claude announcement post on X: https://x.com/claudeai/status/2107522596845822135