Claude Comes to Google Docs, Sheets, and Slides - Sidebar Add-on and Editing Connectors in Beta

On October 6, 2026, Anthropic released Claude for Google Workspace, a sidebar for Google Docs, Sheets, and Slides, as a beta on all paid plans. It also added connectors that let Claude edit Google files from the chat. Edits require approval one by one by default, and ZDR and third-party inference are not supported.

Claude Anthropic Google-Workspace productivity AI-agents enterprise

On October 6, 2026, Anthropic opened a public beta of “Claude for Google Workspace,” which puts Claude inside Google Docs, Sheets, and Slides, for every paid plan¹. Alongside it came beta connectors for Google Docs, Sheets, and Slides that let Claude make new Google files and modify existing ones starting from a chat. In other words, two routes arrived on the same day: one that calls Claude into a Google file, and one that reaches Google files from a Claude conversation.

For organizations whose work lives in Workspace, this adds Claude to the list of AI options available inside Docs, Sheets, and Slides. Whether a team can actually turn it on, however, depends more on how the Workspace admin has configured things than on the Claude subscription. Below we go through how the two routes differ, how edit approval works, and what admins need to check.

The add-on appears as a sidebar next to the open file. Claude reads the document, sheet, or deck in front of you, along with any text, cells, or slides you have highlighted, and changes the file itself¹. A single install from the Marketplace makes it available in all three apps².

Anthropic describes the role in each app as follows¹. In Docs, Claude can correct one sentence or change a heading’s style while leaving the formatting around it intact; for larger rewrites, it shows proposed changes as cards in the sidebar, and the user decides whether to apply or discard each one. In Sheets, it can produce formulas, pivot tables, Sheets’ own charts, and new tabs, and for joins or cleanup work it hands a range to Python and puts the output back in the sheet. In Slides, it creates new slides that follow the deck’s existing layouts and theme, then reviews its own output for overlapping elements, content spilling past the slide edge, and hard-to-read text.

The models, connectors, and skills attached to the signed-in Claude account carry over to the sidebar¹. In Anthropic’s example, a request for a quarterly review deck for a customer can pull that customer’s history out of Salesforce and fetch notes from recent calls via the Google Drive connector.

On the other hand, the add-on can only read the file that is open. According to the help article, aside from content brought in through connectors the user has enabled, Claude has no view of other Drive files, email, or calendar, and the Google permissions it asks for do not extend to Drive, Gmail, or Calendar². Beta limitations listed there include: closing the sidebar halts the task in progress; Firefox is unsupported, so Chrome, Edge, or Safari is required; and Google cuts off any single long operation after six minutes at most. The add-on cannot read Docs comments, and charts it creates in Slides go in as images rather than linked Sheets charts. Conversation history is kept per file in the browser and does not follow you to another device.

Approval per change by default, with some actions always gated

Users pick one of two modes that set how far Claude may go unassisted. Under “Ask before edits,” which is the default, every change surfaces as a card summarizing it, and Claude holds off until the user approves. Switching to “Accept all edits” lets Claude carry the task through to the end, applying changes as it goes¹.

Even in “Accept all edits,” not everything goes through unchecked. As the help article explains, what proceeds without pausing in that mode is routine content editing; anything involving links, outside images, web addresses, or access to the file still requires approval². On top of that, pulling in an image from a URL, adding a link, and removing someone’s access always stop for confirmation whichever mode is selected; these prompts are tagged “Security review” and offer no “always allow” choice. Sharing the file, granting access, changing ownership, and building menus or scheduled automations are refused outright, even on request.

The help article also covers defenses against instructions planted in files. It notes that a file may contain directions written by another person, sometimes concealed in white or very small text, and says Claude handles document content as data, not as commands, flags any hidden text it detects, and is built not to act on a document’s request to share the file, open a link, or reach out to anyone². Even so, it advises checking Claude’s proposals before approving them, particularly for files that came from outside the organization. Running externally received material through “Accept all edits” in bulk would not match that advice.

Google logs Claude’s edits under the user’s own name, and they show up in version history². In Docs, Claude’s changes go in as direct edits, not as suggestions. For anyone who wants to check changes one at a time, the help article recommends keeping the mode on “Ask before edits.”

Connectors for editing Google files from Claude

The second route starts in a Claude conversation. Dropping a Google file’s link into the conversation, or requesting a new Google document, sheet, or deck, brings the file up in a side panel beside the chat on supported setups, where Claude builds or revises it¹. Anthropic says this route fits best when a new file is needed or when the job involves several files. What Claude can reach is bounded by the user’s existing Google sharing permissions.

The help article adds some detail on how to choose³. The user and Claude can both edit one file at once while it is displayed there. To get a Google file, you have to say “Google” explicitly; otherwise a plain request for a “doc” or “deck” results in Claude making a local file.

The connectors and the add-on do not have identical capabilities. The connectors can read and work with Docs comments and suggestions³, while the add-on cannot read comments². In Slides, by contrast, the connectors cannot add charts, and changes to a deck might not appear live³. That side pane works in the web version of Claude opened in Chrome, and in Claude Desktop once its built-in browser is switched on; on Enterprise plans, that built-in browser is turned off by default.

These three connectors are distinct from the existing Google Drive connector, which continues to handle searching for, uploading, and sharing files³.

Workspace admins decide whether it can be installed

The first thing that matters for the add-on is the Workspace configuration. The help article points out that many organizations block Marketplace app installs by default, and that whether an individual can install it is determined by the Workspace admin’s Marketplace settings rather than by the Claude plan². Admins can either put Claude on the allowlist or deploy it through the Google Admin console to everyone or to chosen groups and organizational units. Admins can deploy it this way even if end-user installs stay blocked.

For the connectors, an organization on a Team or Enterprise plan needs its owner or primary owner to switch them on before anyone can use them¹. Workspace may also need to mark Claude as a trusted app, but live editing goes through the “Claude for Google Drive” app that the Drive connector already uses, so a single admin approval covers every file type³.

On Enterprise plans, the add-on is also covered by CMEK (customer-managed encryption keys), the Compliance API, and audit export via OpenTelemetry¹. Per the help article, the Compliance API provides each add-on session’s prompts, responses, and tool calls, along with Claude’s changes to the file².

The help article is equally explicit about what is not available². This beta does not support running inference through third-party platforms such as Amazon Bedrock, Google Cloud Vertex AI, Azure AI Foundry, or an LLM gateway, and signing in with a Claude account is mandatory. The same article states that Claude for Microsoft 365 does support those platforms. Zero Data Retention (ZDR) is unavailable, HIPAA-ready organizations are not covered, and users are told not to use it with protected health information. Custom data retention settings an organization has configured do not apply to the add-on; Anthropic removes inputs and outputs within 30 days, with some exceptions, and says this material is not used to train its models.

That means organizations using Claude via Bedrock or Vertex AI, or relying on ZDR, cannot currently bring in the add-on under the same terms.

Gemini and Claude side by side in Workspace

Google, for its part, has been making changes to the AI features inside Workspace. On September 1 it announced Google Pics, an image generation and editing tool that can be opened from Docs and Slides. On September 30 (US time) it announced that it would introduce reusable custom instructions called “skills” to Gemini and phase out Gems. As an example of an outside tool operating Workspace, Cursor released plugins that work with Google Drive, Gmail, and Google Calendar in August.

Against that backdrop, Claude for Google Workspace becomes one option for using a model from a company other than Google inside Workspace documents. Because Claude’s skills and connectors are available in the sidebar as-is, organizations that have already set up skills and connections to internal systems in Claude are likely to find adoption less work.

Two things are worth checking before trying it: how the Workspace admin restricts Marketplace apps, and whether your contract assumes ZDR or third-party inference. After that, a reasonable order would be to keep “Ask before edits” on while testing with internal documents, settle how externally received files should be handled, and only then decide where “Accept all edits” is appropriate.