Claude Can Now Send Emails in Gmail Without Your Approval
Anthropic's Google Workspace connector already handled Gmail drafts, Calendar, and Drive. Sending was the one action it couldn't take autonomously - until August 18.
Upgraded. Claude Gmail integration gained the one piece it was missing - the ability to actually send. Anthropic's Google Workspace connector already read inboxes, managed calendar events, and handled Drive files. Sending, forwarding, and replying sat outside its reach until August 18. Claude Voice Mode already takes actions in Gmail, Slack, and Notion - the text connector now catches up.
Takeaways
- Claude can now send, reply to, and forward Gmail messages without per-action approval
- Asking for approval before each action remains the default - users disable it manually
- Team and Enterprise owners decide whether members can allow unsupervised sending
- Only available on paid Claude plans - free tier excluded
- Claude's Google Workspace connector already covered Gmail drafts, Calendar, and Drive
Approval is still the default. Anthropic's support page confirms Claude asks before sending, replying, or forwarding unless the user explicitly changes the setting. On Team and Enterprise plans, workspace owners decide whether members can disable per-action approval entirely. Anthropic framed it on X: "Ask Claude to reply to a thread, and it drafts and sends the response. You control when it needs your approval." On personal plans, that control belongs to the user. On team accounts, it belongs to the owner.
Claude Cowork hit web and mobile with one in three sessions flagged as business operations work - email sits at the center of that. Claude Cowork versus ChatGPT Work comes down to depth of real connector action, and Gmail send access is where that depth becomes tangible. 1Password for Claude already lets agents authenticate to sites without exposing credentials - unsupervised Claude Gmail replies are a shorter trust jump than that.
Anthropic did not disclose how many users have disabled the approval requirement, or what guardrails prevent a misfire when an agent misreads a reply target on a long thread. Whether teams trust Claude Gmail to send autonomously - not just draft - is what agentic AI in the workplace has always required: the model's correctness matters, but confidence in it matters more.