Guide
Can Claude read my email? Yes, 4 ways.
Claude can read, search and answer email, and which mailbox it can reach, and what it can do once it is in, depends on which of 4 routes you take. 3 are Anthropic's own. The fourth is the one that reaches everything else, and it is what this site sells, so every claim about it here was measured rather than promised.
- Reading About 13 minutes
- Routes 4, and one is ours
- Vendor pages read 4 October 2026
The short version
- Yes. Claude reads and searches a Gmail or work Microsoft 365 mailbox through Anthropic's own connectors, and reads the message you have open through its Outlook add-in.
- It sends on 2 of those routes. Filing is patchy: the Gmail connector can label, archive and star, the add-in moves and flags the message you have open, and the Microsoft 365 connector only trashes and categorises.
- A custom connector such as Mailbox MCP is how Claude reaches any other mailbox, including a personal outlook.com address or any IMAP host, and how it moves, flags and files what it reads.
- It acts when you ask, or, through Mailbox MCP, in a routine your own Claude runs on a schedule you set. Nothing here reacts to new mail arriving.
The map
The 4 ways Claude can reach a mailbox
"Can Claude read my emails" gets 4 different answers, depending on which Claude product is doing the reading. Most pages that answer it describe one of the 4 as if it were the whole story.
Anthropic runs 3 separate mail integrations, on different plans, with different capabilities. The fourth route is a connector somebody else built, which is what we build. The table is the whole guide in one glance; the sections after it are the evidence for each cell.
| Gmail connector | Microsoft 365 connector | Outlook add-in | Mailbox MCP (custom) | |
|---|---|---|---|---|
| Which mailboxes | Google accounts | Work Microsoft 365, not personal outlook.com | Exchange Online through Microsoft 365 | Gmail, work or personal Microsoft, any IMAP host |
| Claude plan | Stated 2 ways, see below | All plans | Pro, Max, Team, Enterprise | All plans, one connector on Free |
| Reads and searches | Yes | Yes | The open message, plus search | Yes, without marking anything read |
| Sends | Yes, asks first by default | Yes, once an admin enables writes | Never. Drafts land unsent | Yes, and asks first in your client |
| Reads what is inside an attachment | Metadata only; in our test a 1 KB PDF the long way, a 1.6 MB one not at all | Not stated | Word and Excel, on the open message | Yes: PDF, Word, Excel, text and pictures |
| Sends with an attachment | Forwards one; a new file only as base64 written into the chat (our test) | Refused on every write | Not applicable | Yes, up to 20 MB |
| Moves to a folder, flags, marks read | By label: archive, star, mark read; deletes to Trash (tested) | Trash and categories only | Move and flag, via Graph | Yes, all 3 |
Every vendor cell was read from the page cited under it on . Both Microsoft 365 pages were read again on for the row on reading attachments, where neither says a word, and the filing and attachment-sending cells were read again on .
"Not stated" means the vendor's page neither confirms nor denies it. That's a different thing from "no", and the difference matters on the rows that decide whether Claude can run an inbox rather than just look at one.
The full 6-product comparison, with Copilot and Gemini as well, is on the free email connectors compared; this guide stays on Claude.
Route one
Can Claude read my Gmail? Its own connector
Yes, and this is the route most people mean. Under Customize and then Connectors, Claude offers Gmail alongside Google Calendar and Google Drive. Sign in to Google and grant the permissions, and Claude can search the mailbox in natural language, read messages and threads, and draft replies.
It sends as well. Anthropic's help centre, read on , lists the capability in one line, and the second sentence is the one to notice:
Send, reply to, and forward emails from Gmail. By default, Claude asks for your approval before each of these actions.
Anthropic, Use Google Workspace connectors
That "by default" is doing real work. It means the approval is a setting, not a guarantee, and a conversation where it has been switched off will send without showing you the message. It's worth knowing it is a default.
The same page lists what Claude gets about attachments: their metadata and, in Anthropic's words, "not attachment content". It says it again plainly: "Attachment content is not directly accessible through Gmail (metadata only)." Claude can tell you there's a PDF called invoice-0412 on a message, and usually can't tell you what's in it.
When we tested it on 6 October 2026, it read a 1 KB invoice only by pulling the whole raw email into the chat, and couldn't open a 1.6 MB PDF at all; the test is written up here.
So if you're asking can Claude read email attachments, the honest answer through this connector is no. That page shows what a connector that does read them returns, file type by file type.
The plan question has 2 answers on Anthropic's own site, and both are quoted on the comparison page. The help centre says the Google Workspace connectors are available for all users; the developer documentation names Pro, Max, Team and Enterprise. Your own Connectors screen is the only answer that counts.
Besides sending, the only change the help page names is "labels". That's more than it sounds, because Gmail files by label: archiving, marking read and starring are all labels underneath.
Moving to a folder and flagging, in the Outlook sense, aren't Gmail ideas at all. The section on managing an inbox below shows what the connector's published tools can do.
Routes 2 and 3
Can Claude use Outlook? 2 Microsoft routes, and one mailbox neither reaches
Yes, with a sharp condition on which Outlook. Anthropic's Microsoft 365 connector is available on every Claude plan including Free, and it connects a work or school mailbox. In the set-up article's own words, it "requires a Microsoft Entra tenant tied to a Microsoft Business plan", and "Personal Microsoft accounts (such as @outlook.com or @hotmail.com addresses) can't be used to connect."
So the Outlook most people have at home, the free one on outlook.com, is exactly the one Anthropic's connector refuses. Claude with a Hotmail or Outlook.com address takes a custom connector, and that guide shows the one setting to turn on first.
On a work mailbox it reads everything, and an administrator switches writing on. Anthropic publishes the connector's full tool list, the most checkable capability evidence any vendor gives, and the list sends, forwards, drafts, trashes and applies categories.
Its security guide, read the same day, also states that every write tool rejects a message carrying an attachment, and that sent mail carries a marker:
Emails Claude sends include an attribution header identifying them as agent-initiated.
Anthropic, Microsoft 365 connector security guide
Read that beside the founding rule of this product: a recipient shouldn't be able to tell a message was handled by an assistant. That's the clearest statement of the difference between the 2 designs.
Anthropic's connector labels the mail; ours threads it, files one copy in Sent and leaves nothing else behind. Neither is wrong. They're answers to different questions, and it's better to know which one you're buying.
The third route is Claude for Outlook, an add-in that puts Claude inside the Outlook window on Pro, Max, Team and Enterprise. It reads the open message and its Word and Excel attachments, searches the mailbox, and, unusually, moves and flags messages through Microsoft Graph.
What it never does is send, in the words of its own documentation, read the same day:
Claude never sends mail or invites on its own. The add-in does not request the Mail.Send permission. Every draft lands unsent in Outlook's compose or appointment form, and you click Send.
Anthropic, Use Claude for Outlook
That sentence matters because it names the permission rather than the intention. A product that hasn't been granted Mail.Send can't send, whatever a prompt or a hidden instruction in an incoming message asks it to.
Mailbox MCP gives you the same kind of guarantee on any mailbox, by level. Connect it at Draft and file and Claude drafts and files but can't send. A draft it saves is sent by you, from your own mail client.
A personal outlook.com mailbox connects here with the same single Microsoft sign-in, on the free tier, with no card.
The gap
Can Claude manage my inbox? What Anthropic's 3 routes file, and what they don't
Managing an inbox means filing, not just finding. Put the 3 routes side by side and they file very differently, and the differences are in the vendors' own documentation rather than in anybody's opinion of it.
Gmail is the exception now. Anthropic's listing for
its Gmail connector names Google's server and publishes its tool names,
label_message and unlabel_message among them.
Google's own reference says those take system labels such as INBOX,
UNREAD and STARRED, which is how Gmail archives, marks read and stars.
Neither company uses those words for it, so treat it as filing by label.
The Microsoft 365 connector is where the gap stays. Its published tools trash a conversation, set categories and even create an inbox rule, and there's still no tool to move a message to a folder, flag it or mark it read. So on a work Outlook, Claude can find the 40 messages that belong in a project folder and can't put them there.
The add-in moves and flags, but only as part of an action you take on the message you have open in Outlook.
And on the question underneath all of this, whether Claude reading a message through any of these routes marks it as read, no vendor page we could find says anything at all. That silence is the difference between a measurement and a promise. It decides whether a triage pass you abandon halfway leaves your unread count meaning anything.
And none of them reaches a mailbox that isn't the vendor's own. Google accounts on one route, business Microsoft tenants on the other 2.
A Fastmail, iCloud, Zoho or cPanel mailbox, or the outlook.com address you've had since 2009, can't be connected to Claude through anything Anthropic ships. That's not a criticism of Anthropic, who never claimed otherwise. It's the reason the fourth route exists.
The 6 rows below are what "manage my inbox" turns on, measured on the fourth route against a live mailbox of 26,930 messages rather than inferred from an API response. They are the rows the vendor pages leave blank.
- Reading a message Does not mark it read
- Moving a message Flags and original date kept
- Deleting a message Moves to Trash, never expunges
- Replies In-Reply-To and References carried
- Copies filed per send 1, no duplicate
- Attachment bytes SHA-256 identical on arrival
Route 4
The fourth way: a custom connector, on the Claude plan you already have
Claude can connect to servers Anthropic didn't build, through the same Connectors screen, by pasting a URL. Anthropic calls these custom connectors, and the protocol they speak is MCP: the part of this product's name that nobody searching "can Claude read my email" has ever typed.
What matters here is the plan gate, read the same day, because it's the opposite of what people assume:
Custom connectors using remote MCP are available on Claude, Cowork, and Claude Desktop for users on free, Pro, Max, Team, and Enterprise plans. Free users are limited to one custom connector.
Anthropic help centre, custom connectors
So a free Claude account can connect one mailbox this way, and a paid one as many as it likes. Running multiple email accounts in Claude side by side, named so it picks the right one, has a guide of its own.
The same page carries an example of what an organisation can restrict a connector to, and it happens to be our product described in reverse: "Allow Claude to search and summarize email, but prevent it from sending messages." The control sits with the Claude side, on top of whatever the connector offers.
Mailbox MCP is one such connector. You connect the mailbox to it (Gmail with an app password, Outlook or Microsoft 365 with one Microsoft sign-in, or any IMAP host with the settings you already have), then add one URL to Claude.
Claude is then handed 37 email tools, each named on the tools page with what it refuses, and the mailbox behaves as the ledger above says. The mailboxes this product was built against are the author's own, and they're connected to Claude through it as this is written.
On a paid Claude plan, your Claude can also run one of our routines on a schedule you set, such as a morning brief at 07:00 on weekdays.
It reads what's inside an attachment, too. Ask what an attached invoice comes to and Claude gets the PDF as text, page by page. A Word document keeps its tables, a spreadsheet comes back as rows with the formulas already worked out, and a slide deck slide by slide. A photo, or a scanned letter, comes back as pictures Claude can look at.
The file itself never passes through the conversation as bytes, only what it says. Each attachment also comes with a short-lived download link, so Claude can hand you the file, or attach it to a new message without opening it. How AI email attachments work has the rest.
What does it cost on top of Claude? The answer is stated once, everywhere: 5 calls a day on any mailbox for nothing, and £2.92 a month per mailbox, paid annually at £34.99 + VAT, for 1,000. There's no second AI subscription in that: it uses the Claude you already pay for, or the free one.
Before you start
When you need Mailbox MCP in Claude, and what it doesn't do
You want the fourth route when:
- the mailbox is on a provider Anthropic doesn't connect, including a personal outlook.com address
- you want Claude to file, flag and mark as well as read
- a reply has to land inside the thread it belongs to, and file one copy in Sent
- an attachment has to go with the message
- you'd rather reading a message left it unread
Each of those is a row in the table at the top, and each is the reason a particular customer came here.
What Mailbox MCP doesn't do. One connector opens one mailbox, so 3 mailboxes are 3 connectors, each signed in on its own and each with its own permission level.
Gmail connects with an app password, which needs 2-Step Verification backed by a phone or an authenticator app. It never touches your Google password, and you can withdraw it on its own.
And nothing runs without your AI. There are no rules or new-mail alerts on our server: Claude does the work, in a conversation or in a routine it runs on a schedule you set.
If you have already hit one of those rows, the free tier is enough to check the fourth route on a mailbox you do not mind.
Limits
What this guide cannot tell you
-
Tomorrow's Anthropic
Whether Anthropic's connectors still do what its pages said on 11 September 2026. They change monthly, the plan gates move, and the dates on every quotation here are there so you can tell how old a claim is.
-
The unread flag, on their routes
Whether reading through Anthropic's own connectors marks a message read. No vendor page says. We measured our own route and say so; we cannot measure theirs for you.
-
Anything that reacts to new mail
Whether Claude can watch an inbox and act the moment mail arrives. Nothing on this server reacts to new mail, and the server schedules nothing itself. What runs while you are away is a scheduled task in your own Claude, on a paid plan, such as one of our routines, at the time you chose.
-
Your employer's rules
Whether your organisation permits any of it. On a work mailbox an administrator decides, and on the Microsoft 365 connector Anthropic gives them the switch for writes.
Can Claude read my emails?
Yes. Claude reads and searches a Gmail mailbox through Anthropic's own Gmail connector, a work Microsoft 365 mailbox through its Microsoft 365 connector, and the message you have open in Outlook through the Claude for Outlook add-in. It reads a mailbox on any other provider, including a personal outlook.com address or any IMAP host, through a custom connector such as Mailbox MCP. It reads when you ask, or in a routine your own Claude runs on a schedule.
Can Claude send emails for me?
Yes, on 3 of the 4 routes. Anthropic's Gmail connector sends, replies and forwards, asking for your approval first by default. The Microsoft 365 connector sends once an administrator enables its write tools, and rejects any message with an attachment. The Outlook add-in never sends. Through a custom connector like this one, Claude sends from any connected mailbox with attachments up to 20 MB, files one copy in Sent, and threads the reply into the original conversation.
Can Claude use Outlook?
Yes, and which route depends on which Outlook you mean. A work or school Microsoft 365 mailbox connects through Anthropic's Microsoft 365 connector on any Claude plan, or through the Outlook add-in on a paid plan, or through a custom connector. A personal outlook.com or hotmail.com address is refused by Anthropic's own connector, in its own words, so the only route to one of those is a custom connector: Mailbox MCP connects it with a single Microsoft sign-in.
Can Claude manage my inbox?
Yes, and how far depends on the route. Anthropic's Gmail connector files by label, so it can archive, star and mark read; its Microsoft 365 connector only trashes and sets categories, with no tool to move, flag or mark a message read. A custom connector like this one gives Claude 37 email tools, including move, flag, mark read, archive and junk. Through it your own Claude can also run a routine on a schedule. Nothing reacts to new mail arriving.
Which Claude plan do I need to connect my email?
For a custom connector such as Mailbox MCP, any plan: Anthropic says custom connectors using remote MCP are on free, Pro, Max, Team and Enterprise, with free accounts limited to one. Its Microsoft 365 connector is on all plans too. Its Gmail connector is for all users in the help centre and Pro, Max, Team and Enterprise in the developer documentation; the Outlook add-in is Pro, Max, Team and Enterprise. Check your own Connectors screen, because the gates move.
Does Claude reading a message mark it as read?
Through Mailbox MCP, no: the engine fetches with IMAP's BODY.PEEK, which leaves the unread flag exactly as it was, and marking a message read is a separate tool that runs only when you ask. Through Anthropic's own connectors, nobody knows, because no vendor page we could find says either way. That silence is the difference between a measurement and a promise, and it is worth checking on a message you do not mind before you trust either.
Can Claude read my email attachments?
Anthropic's Gmail connector gives Claude attachment metadata only, in its own words "not attachment content". The Outlook add-in reads Word and Excel attachments on the open message. Mailbox MCP reads them on any mailbox, however many are on the message: a PDF as text, a Word document with its tables kept, a spreadsheet as rows with the formulas worked out, a deck slide by slide, and a photo or scan as pictures. None of it marks the message read.
How do I stop AI from reading my emails?
Take the connection away and the reading stops. In Claude, open Customize, then Connectors, and disconnect the mail connector, or turn it off for one chat from the + menu. For Mailbox MCP you can also revoke the connection in our control panel, which ends it at our end whatever Claude does. To keep the help but stop it changing anything, choose the Read only level for it there.
Attribution
Sources
- Anthropic: Use Google Workspace connectors The Gmail connector's capability list, the send-with-approval sentence, and the attachment metadata limit. Read 11 September 2026, and again on 4 October 2026 for "metadata only".
- Anthropic: the Gmail connector listing Google's connector server and the published tool names, label_message, unlabel_message and trash_message among them. Read 4 October 2026.
- Google: the unlabel_message tool System labels such as INBOX, UNREAD and STARRED, which is how archiving, marking read and starring are done. Google's create_draft page in the same reference says drafts with attachments aren't supported yet. Read 4 October 2026.
- Anthropic: Set up the Microsoft 365 connector Availability on all plans, the Entra tenant requirement, and the refusal of personal outlook.com and hotmail.com accounts. Read 11 September 2026.
- Anthropic: Microsoft 365 connector security guide The published tool list, the admin gate on write tools, the attachment refusal on every write, and the attribution header. Read 11 September 2026, again on 12 September for whether it says anything about reading an attachment's contents (it does not), and on 1 October for the filing tools: trash, categories and an inbox rule, and no move, flag or mark read.
- Anthropic: Use Claude for Outlook Plans, supported Outlook versions, move and flag through Microsoft Graph, and the statement that the add-in never sends. Read 11 September 2026.
- Anthropic: Use connectors to extend Claude's capabilities Custom connectors on every plan including Free, the one-connector limit on Free, and the search-but-do-not-send example. Read 11 September 2026.
- Our comparison of the free Gmail and Outlook connectors The capability table for 7 free connectors, Claude's 3 among them, with every cell sourced and dated, and the vendor self-contradictions in full.
- Our security page Every claim about this product's own behaviour, verified against the engine source rather than assumed.
Keep going
Read next
Connect the mailbox Claude cannot reach on its own.
A personal Outlook, an IMAP host, or a Gmail you want filed rather than searched. 5 calls a day, no card, and one URL in Claude's Connectors screen.