Claude Can Now Send Emails in Gmail Without Per-Action Approval
Anthropic has crossed a notable threshold in its agentic product rollout: Claude can now send emails in Gmail without asking users to approve each outbound message. The capability arrives alongside the broader rollout of Claude Cowork to all paid plan tiers on mobile.
Until this update, Claude’s Gmail integration covered reading, searching, drafting, and organizing. The ability to actually send — to dispatch a message on a user’s behalf with no per-action prompt — was a deliberate holdout. It no longer is.
What Changed
The distinction matters because email is asymmetric: reading a message carries no external consequence, but sending one reaches another person without the possibility of withdrawal. Claude now operates in that send-capable state by default for users who have connected Gmail to Claude Cowork.
Anthropic describes the capability as part of Claude Cowork’s expansion from Mac to iPhone and iPad, which became available across all paid plans this week. The Gmail send capability appears to have shipped alongside this rollout rather than as a separate feature announcement.
The Trust Shift
The product decision encodes a specific stance on AI agent trust: that a general-purpose assistant with calendar, email-read, and task context is reliable enough to send outbound communications without per-message human confirmation.
This is different from existing AI email tools that operate in “suggest and send” mode — where the user clicks a final send button — or draft-only modes that keep Claude in a read-only position. Operating without per-action approval puts Claude in the same category as a human employee with inbox access.
The shift follows a pattern in Anthropic’s product decisions: Claude Code moved to Auto Mode by default on August 14, removing the step that required human confirmation before each code execution. Gmail send authority is the same logic applied to a communication surface.
What It Means for Agentic Deployment
Autonomous email send is the first high-consequence unilateral action Anthropic has enabled in a consumer-facing Claude product at scale. Previous autonomous actions — file edits, code execution, calendar reads — either operate within a sandbox or involve reading rather than writing to external systems.
The Gmail integration is now read-write-send. That is a meaningfully different threat surface than read-write-draft. The open question is what other communication channels — Slack, WhatsApp, calendar invites — follow the same authorization model as Gmail does today.