ExplainerUpdated October 17, 2026

Can ChatGPT send emails for you? And can Claude?

Yes. ChatGPT can send email once you connect a mail account such as Gmail or Outlook and that app offers a send action. OpenAI treats sending an email as an action that may need extra review, so depending on your permission setting ChatGPT asks first or refuses[1]. Claude can send, reply and forward through its Gmail connector, and it asks for your approval before each of those actions by default[7].

In ChatGPT, whether a send needs your confirmation depends on the app, the connected account and your permission settings. How much control you keep beyond that prompt depends on the setting you pick, which accounts you connect and whether an agent is supposed to work while you are away.

Updated: Fact-checked: About 7 min readAs Markdown

Contents
  1. How sending works in ChatGPT
  2. Can Claude send emails?
  3. Where the built-in options run out
  4. What happens to the content of your mail
  5. Practical ways to stay in control
  6. When a dedicated agent inbox makes sense
  7. Frequently asked questions
  8. Sources

01How sending works in ChatGPT

ChatGPT sends mail through connected apps, and a permission setting decides when it stops to ask. On an eligible personal account you set the default under Settings → Plugins → Permissions; if your Settings show Apps instead of Plugins, check the controls offered there. You can also give a single connected account its own setting, different from other accounts on the same app. Depending on your account and app, there are up to four options[1]:

  • Always ask: ChatGPT asks before reading app data and before making changes.
  • Allow read actions: it reads freely but asks before changes such as a send.
  • Allow low-risk actions: low-risk actions run on their own; higher-risk ones need confirmation or are denied.
  • Allow all actions: supported actions run without further prompts. OpenAI calls this elevated risk and does not offer it in the account-wide or workspace-wide default.

When an action needs approval, ChatGPT shows a card naming the app, the account and the proposed step. You can deny it, allow it once, allow low-risk actions for that account, or, for an eligible personally connected account, always allow. Members of managed workspaces don't get the persistent option[1].

What the email apps can do differs by provider. For Outlook, OpenAI lists the Microsoft permission Mail.Send and describes sending plain-text email from or on behalf of shared and delegated mailboxes[2]. For Gmail, ChatGPT requests the Google scope gmail.modify[3], which Google describes as permission to read, compose and send email, short of permanent deletion[4]. That scope alone doesn't prove a send tool is available on your account: the OpenAI articles we checked describe a concrete send action for Outlook, but not for Gmail. Which actions you actually see still depends on the app, your plan, your region and your workspace settings.

02Can Claude send emails?

Yes. Claude's Gmail connector searches and reads mail, writes drafts, sends, replies, forwards, and manages labels and threads[7]. Claude asks before sending, replying or forwarding unless an owner on a Team or Enterprise plan has allowed members to skip the prompt.

The Microsoft 365 connector is available on every Claude plan but only for work accounts in a Microsoft Entra tenant. Personal addresses at outlook.com, hotmail.com or live.com are not supported[8]. Claude can send from Outlook there only if an admin has turned on write tools. Messages Claude sends carry a header that marks them as agent-initiated, and the write tools can't add attachments yet.

03Where the built-in options run out

For a single Gmail or work Outlook account, the built-in connectors cover a lot. The friction shows up in a few predictable places:

  • One account per connection. Claude works with the one Google account you connected[7], and a request for multiple Gmail accounts has been open since March 2026[9]. ChatGPT supports extra accounts for some apps, though not all[5]; for Codex, users have asked for multiple named Gmail accounts in an open issue[12].
  • Attachments. Claude sees Gmail attachment metadata, not the file contents[7].
  • Success messages you can't trust blindly. One open report describes a scheduled Claude routine that reported success although its Gmail send_message call did nothing[10]. For Outlook, OpenAI tells users to check the mailbox and sent items before repeating a send whose outcome is unclear[2].
  • Forward means send. Another open request asks for Claude's forward action to create a draft instead of sending right away[11].
  • Approval is per action. ChatGPT lets each connected account have its own permission level[1], but neither provider's help pages document a daily sending cap you set yourself.

04What happens to the content of your mail

The Gmail and Outlook connections described here get read access to your mailbox along with the ability to send. The provider's sign-in screen shows which permissions a connection actually requests, so read it before you agree.

Anthropic says it does not train its models on Gmail, Drive or Calendar connector data, and retrieved data stays with the chat it was used in until you delete that chat[7]. It names one exception for Free, Pro and Max: if you allow your chats to be used for training, content you copy from these services and Claude responses that include specific information from them may be used to improve its models.

OpenAI says it does not train its generalized models on data accessed directly from connected Google apps. The exceptions are conversations you submit as feedback, content you copy, paste or upload yourself, and content included in ChatGPT's response. With "Improve the model for everyone" switched off, OpenAI doesn't use that app data either[3]. For other connected apps on Free, Plus, Go and Pro, the same setting decides[6].

05Practical ways to stay in control

These habits help with any assistant:

  1. Start with confirmation on. Keep Claude's default prompt and set ChatGPT to Always ask or Allow read actions.
  2. Ask for drafts. Phrase requests as "draft a reply, but don't send it". Anthropic uses exactly that wording as an example for its Microsoft 365 connector[8].
  3. Read the approval card. Check recipient, account and text before you approve; an approval covers the step shown on the card.
  4. Check Sent before retrying. After an error or a timeout, look in the sent folder first.
  5. Treat incoming mail as untrusted text. A message can carry hidden instructions. Connect only the accounts an agent needs and add send rights after reading and drafting work well.

06When a dedicated agent inbox makes sense

A separate tool earns its place when you juggle several addresses, when an agent should run on a schedule without you, or when several agents should follow the same sending rules. Command Mail is our product, and it is built for those cases[14].

Control in Command Mail comes from individual permissions you approve when you connect an agent[15]. With mail:draft the agent saves drafts and never sends. With mail:send-approved it sends exactly the draft version a person approved, once; if the text changes afterwards, it needs a new approval. Only mail:send lets an agent send on its own, with a unique send key per email, a daily limit of 50 by default and at most 20 recipients per message. Sending without approval starts switched off.

For sorting, filing and rules, each mailbox also has an autonomy level: read, suggest, rules or auto, and new mailboxes start on auto. That level does not limit sending. Sending permissions apply to the whole account, so an agent with mail:send can send from a mailbox set to read. Don't give mail:send to an agent that should only sort.

Command Mail treats message content as untrusted data and flags suspected prompt injection. If an agent uses the reply tool on a message whose headers mark it as automated, such as Auto-Submitted, Precedence: bulk or List-Id, Command Mail refuses the reply. The same rules apply whether the agent is Claude, ChatGPT, Codex, Cursor or another MCP client.

Mail reaches Command Mail as forwarded copies, and changes don't sync back to the original mailbox. The Free plan doesn't send; sending starts with Pro at €19 a month as a founder price for the first 100 paying Pro accounts and €29 after that[15]. In Claude you connect Command Mail from the connectors directory. In ChatGPT our directory listing is still under review, so for now you add it under Plugins as a custom MCP server, without turning on developer mode first. OpenAI allows that on the web only: Pro with read and fetch actions, and Business, Enterprise and Edu with full support including write actions, subject to roles and workspace settings[13]. Sending from ChatGPT therefore needs one of those three plans.

The For agents page and the documentation cover permissions and connection options in detail. To give one agent all your addresses, follow Use multiple email accounts with an AI agent. For the built-in route, see Connect Gmail to Claude. Other ways to give an agent email access are compared in Best email MCP servers for Gmail, Outlook and IMAP.

You can start free and connect your agent with exactly the permissions you want it to have.

Frequently asked questions

Will ChatGPT send an email without asking me?

Only if you let it. On Always ask or Allow read actions, ChatGPT asks before any change. Allow all actions lets supported actions run without another prompt; OpenAI labels it elevated risk and keeps it out of the account-wide default selector. Each connected account can have its own setting.

Does Claude ask before it sends an email?

Yes. With the Gmail connector, Claude asks for approval before each send, reply and forward by default. On Team and Enterprise plans, owners decide whether members may switch that prompt off.

Can Claude use my personal Outlook.com address?

Not through the Microsoft 365 connector. Anthropic requires a work account in a Microsoft Entra tenant and does not accept @outlook.com, @hotmail.com or @live.com addresses.

Can ChatGPT or Claude attach files when they send?

Claude's Microsoft 365 write tools can't send, forward or draft emails with attachments yet, and its Gmail connector reads attachment metadata only. OpenAI's Outlook article covers reading attachments in your own mailbox but says nothing about attaching files to outgoing mail.

How do I connect Command Mail to ChatGPT today?

Our ChatGPT directory listing is still under review. Until it is live, you add Command Mail as a custom MCP server under Plugins, with no need to turn on developer mode first. OpenAI allows this on the web only: Pro gets read and fetch actions, while Business, Enterprise and Edu get full support including write actions such as sending, subject to roles and workspace settings. In Claude, Command Mail is in the connectors directory.

Sources

  1. OpenAI Help Center: Managing app permissions in ChatGPT help.openai.com, accessed
  2. OpenAI Help Center: Outlook Email and Calendar apps in ChatGPT help.openai.com, accessed
  3. OpenAI Help Center: Google app data controls FAQ help.openai.com, accessed
  4. Google for Developers: Choose Gmail API scopes developers.google.com, accessed
  5. OpenAI Help Center: Connecting and managing app accounts in ChatGPT help.openai.com, accessed
  6. OpenAI Help Center: Connected apps in ChatGPT help.openai.com, accessed
  7. Claude Help Center: Use Google Workspace connectors support.claude.com, accessed
  8. Claude Help Center: Connect to Microsoft 365 support.claude.com, accessed
  9. GitHub anthropics/claude-code #36024: Support multiple Gmail accounts in MCP integration github.com, accessed
  10. GitHub anthropics/claude-code #97895: Gmail send_message silently no-ops but run reports SUCCEEDED github.com, accessed
  11. GitHub anthropics/claude-code #100177: Gmail connector, let forward create a draft instead of sending github.com, accessed
  12. GitHub openai/codex #30418: Gmail connector should support multiple named accounts github.com, accessed
  13. OpenAI Help Center: Connecting custom MCP servers in ChatGPT help.openai.com, accessed
  14. Command Mail: For agents commandmail.app, accessed
  15. Command Mail: llms.txt with plans and permissions commandmail.app, accessed
EN
DE