Claude Gmail and Drive Connectors Get Write Actions — With Approval by Default

Claude can now send your email and trash your Drive files — but only after you approve each action, unless your admin says otherwise. On August 18, 2026, Anthropic expanded its Google Workspace connectors so Claude can send, reply to, and forward Gmail and share, move, and trash Google Drive files on your behalf. The write actions ship with human-in-the-loop approval turned on by default; on Team and Enterprise plans, owners decide whether members can skip per-action confirmation.

The same day, The Verge reported that Claude Cowork rolled out to all paid accounts on mobile and web — widening the surface where users can monitor agentic workspace automation, not just the desktop client.

For builders wiring agents to real inboxes and file systems, this is less a novelty feature and more a permission-model stress test: the Gmail connector page lists Claude Code and the Claude API alongside chat, which means write-capable email access can inherit into coding sessions the same way read-only connectors already do — a pattern explainx.ai flagged in the Claude connectors guide when Anthropic confirmed chat-configured connectors flow into Claude Code automatically.

TL;DR

What Anthropic actually added

Google Workspace connectors for Gmail, Calendar, and Drive have been available since early 2026 — read-heavy workflows like searching mail, pulling Doc text, and checking calendar availability were already in production. The August 18 update closes the gap on destructive and outbound actions that turn Claude from a research assistant into an actor inside your Google account.

Gmail: from read-only inbox to outbound mail

Per Anthropic's Help Center, Gmail now supports:

• Search and read emails with natural-language queries
• Draft emails with formatting and context
• Send, reply to, and forward — the new write surface
• Label and thread management, draft listing, attachment metadata (content not directly accessible)

Anthropic's Gmail connector page frames the capability plainly: Claude can "draft, send, reply to, and forward emails, and asks for your approval by default before anything goes out." Example prompts Anthropic publishes include "Draft a response to the vendor's pricing proposal… then send it" and "Forward the signed contract to legal with a short summary."

The Help Center repeats the approval rule three times across Gmail permissions, Drive actions, and FAQ — a signal that Anthropic expects pushback on unsupervised email sending.

Google Drive: share, move, trash

Drive write actions expanded beyond creating folders and uploading files:

• Share files with new collaborators
• Move files between folders
• Trash files

Read capabilities — Docs text extraction, Sheets/Slides/PDF preview, metadata lookup, saving Claude-generated files back to Drive — were already documented. The new actions are the ones with no undo in the same sense as reverting a chat message: a shared link persists, a trashed file enters Google's deletion pipeline.

Again: approval required by default for share, move, and trash.

The approval model — and who can turn it off

Anthropic's default is explicit per-action confirmation before Gmail sends or Drive mutations execute. That aligns with how most security-conscious teams want agent tooling to behave when the blast radius includes external recipients and shared permissions.

The org-level override matters for builders:

This is materially different from a "trust the model" posture. It is also different from the worst-case reading some coverage emphasized — headlines about sending "without your approval" refer to the admin-configurable opt-out, not the consumer default. Anthropic's own FAQ: "Claude asks for your approval by default before each of these actions."

For teams evaluating headless SaaS and per-interaction agent pricing, Google's first-party connector surface is instructive: a genuinely complete agent interface (read and write) with a documented approval gate and admin policy layer — closer to the "full parity with guardrails" model than a demo MCP server with three read tools.

Why builders should care about Claude Code and API inheritance

The Gmail connector directory listing is easy to miss and hard to unsee once you have shipped anything that combines inbox access with code execution:

Used in: Claude desktop app · Claude mobile app · Claude Code · Claude API

That means a Gmail connector configured for chat is not siloed to chat. explainx.ai's connectors guide already documented that connectors set up on claude.ai propagate to Claude Code without separate setup — convenient for workflow continuity, dangerous if you treat "I connected Gmail for summarization" as a read-only grant while Claude Code runs shell commands in the same account.

Practical checklist before enabling write-capable Gmail or Drive on any surface:

1. Map which harnesses share the OAuth grant — chat, Cowork, Claude Code, API clients under the same account.
2. Disable write tools you do not need per conversation via the Connectors menu; do not leave send/trash enabled "just in case."
3. Treat prompt injection as in-scope — an email body or shared Doc comment can instruct Claude to forward sensitive threads or share files broadly; approval dialogs are your last line of defense unless an admin disabled them.
4. On Team/Enterprise, audit org connector policy before rolling out to engineers with Claude Code — the owner toggle for skipping approval is a policy decision, not a personal preference.
5. Compare to Cursor's IDE plugins if your team splits time between Claude and Cursor Google Workspace plugins — two OAuth grants, two approval models, one inbox.

Claude Cowork on mobile and web — same day, different story

The Verge's August 18 coverage paired the connector expansion with a separate rollout: Claude Cowork is now available to all paid accounts on mobile and web, not only the desktop Cowork client. Cowork is Anthropic's shared, server-side workspace where Claude organizes files, runs multi-step tasks, and interacts with connected apps — explainx.ai has covered its safety surface in the Claude Cowork safety guide and Cowork use cases roundup.

Mobile and web access does not change what Cowork can do with connectors; it changes where you watch it happen. For anyone running long-horizon automations over Gmail or Drive, that visibility matters — approval prompts and file mutations are easier to miss on a phone notification than on a desktop session you are actively supervising.

What people are asking

"Can Claude send email without me seeing it?"

Not by default. Anthropic requires approval before each send, reply, and forward unless a Team or Enterprise owner has allowed members to opt into running without per-action confirmation. Even then, you still initiate the task — Claude does not autonomously monitor your inbox and send mail on a schedule through this connector alone.

"Is this on the free plan?"

Google Workspace connectors (Gmail, Calendar, Drive) are listed as available on Claude and Claude Desktop broadly; Anthropic's connector directory and Help Center focus on authenticated Google accounts rather than plan-gating the read/write matrix in the August update. Cowork mobile/web expansion specifically targets paid accounts, per The Verge.

"How does this compare to ChatGPT's Google integrations?"

Both ecosystems are racing toward the same end state: models that read and write inside Workspace. explainx.ai covered ChatGPT's Google Docs/Sheets/Slides integration earlier in August. The differentiator for Claude in this release is not raw capability but documented approval defaults and enterprise owner controls — the knobs security teams ask for before granting an agent outbound email.

"Should I connect Gmail to Claude Code?"

Only if you have reviewed the combined blast radius: inbox write access plus local file and shell access in the same agent session. If you need read-only context, disable send/reply/forward tools explicitly rather than relying on model discipline.

The takeaway

Anthropic's August 18 update makes Claude a full participant in Gmail and Drive workflows — not just a reader. The product decision that matters for builders is not whether send/trash exist (competitors were already moving there) but that approval is the documented default and Enterprise owners hold the key to relaxing it org-wide.

Pair that with Gmail's availability on Claude Code and the Claude API, and the security story becomes: your connector grant is a cross-harness permission, not a chat-only feature. Wire it accordingly.

Related on explainx.ai

• How to Use Claude Connectors and MCP Servers: Complete Guide · Top 10 Claude Connectors
• Cursor Google Workspace Plugins: Gmail, Drive, Calendar in the IDE
• Headless SaaS for AI Agents: Charge Per Interaction
• What is MCP? Model Context Protocol Guide
• How to Use Claude Cowork Safely
• Claude Code MCP Servers: Connect Any Tool
• ChatGPT Google Docs, Sheets, Slides Integration
• ChatGPT for Teens — OpenAI's August 2026 safety launch
• Claude protein design results

Official sources: Anthropic — Use Google Workspace connectors · Gmail connector · The Verge — Aug 18, 2026

Connector capabilities, approval defaults, and Cowork availability reflect Anthropic's documentation and The Verge's reporting as of August 18–19, 2026, and may change as Google API quotas, admin policies, or connector surfaces evolve.