Anthropic has announced that Claude can now send email in Gmail and manage files in Google Drive. Ask it to reply to an email thread, and it will draft the response and send it; which actions require a human go-ahead is left for the user to configure. The feature rolls out through the connectors menu, and the official announcement says it is available across the paid plans.
"Claude can now send emails in Gmail and manage files in Google Drive."
Until now, both connectors were essentially read-only. Claude could search a mailbox, pull out the to-dos and the messages that mattered, and summarize them, but sending the reply was still a manual step. Drive worked the same way: search and read were fine, writing required a workaround.
What it can do, and cannot
According to the list published in the Help Center, all three connectors were upgraded together this round:
- Gmail: search and read email, draft, send, reply to and forward messages, organize with labels and threads, read message metadata, and list existing drafts. It cannot read attachment content, and advanced filter rules are not supported.
- Google Drive: search documents, read Sheets, Slides, PDFs, images and Office files, upload files in any format with optional conversion to Google formats, create folders and save generated files into them, view permissions and revision history, and share, move or trash files. It cannot process images embedded in documents, and it cannot read comments or suggested edits inside Docs.
- Google Calendar: view schedules across shared calendars, create, modify and delete events, find common free slots among attendees, handle invitations, and arrange recurring meetings.
The boundary this list draws is fairly clear: sending and filing can be handed off, but rewriting something someone else already wrote still cannot. Editing body text in Docs, along with its comments and suggested-edits layer, is outside the permission scope - the collaborative-writing layer has not been opened up.
Who holds the switch
The default setting requires explicit user approval before sending, replying to, or forwarding anything. On Team and Enterprise plans, whether members can turn that confirmation step off is up to the administrator.
That design moves the risk from the product to the configuration. If the model misreads a message's tone or picks the wrong recipient, who is responsible depends on who switched on unattended execution and who they granted it to in the first place. Corporate IT will likely slow this down with process - the reason tools like Copilot have crept into large enterprises slowly over the past few years usually is not that the model is not good enough, but whether audit logs can be kept and permissions can be revoked with one click.
How it differs from Google's own tools
Gemini is a native component inside Gmail; its permissions follow the Workspace account, administrators manage it from the same console, and revoking access and deactivating the account are the same action. Claude, by contrast, works through an OAuth connector, so authorization, revocation and auditing are split across two systems. For an individual user, it is a one-click setup; for a team with strict compliance requirements, that path requires extra inventory work.
Coverage is the other difference. Google's own assistant serves only Google's products, while Claude's connector menu hooks into other companies' services at the same time, stringing email, files and code repositories into one chain - that is the position from which it can hold its own against a native assistant. The cost is that every additional connector widens the permission surface a little more.
What this means for Chinese-language readers
Gmail and Drive see limited use inside China, so the direct audience for this update is teams with overseas collaboration needs. The reference value lies elsewhere: once "read" is upgraded to "write," the standard for judging an office assistant changes. An inaccurate summary can be fixed with a glance; an email sent to the wrong person is far more expensive to walk back.
The handful of domestic vendors building AI into office suites face the same set of questions: who approves an action, how it gets rolled back when it goes wrong, and who is accountable when it does. Matching the feature set is not the hard part - building the permissions and audit trail solidly enough is, and that work never shows up in a launch demo, even though it decides whether an enterprise customer will dare flip on unattended execution.
Sources: Anthropic official announcement, Claude Help Center's Google Workspace connector documentation, CocoLoop, Android Authority. Capability lists and default approval settings follow the Help Center entries.