If you want to know how to connect Claude to Gmail and Google Calendar, the click itself takes about ninety seconds. Open Claude, go to Customize, then Connectors, find Gmail, click Connect, sign in to Google. Done. Then do it again for Google Calendar.
That is not the hard part.
The hard part is the permission screen you click through on the way, and what you let the AI do once it is inside. Most owners never read that screen. They hand a piece of software the ability to read, re-label, send and permanently delete every email they have ever received, and then they ask it to "tidy the inbox." I have watched what happens next, and it is why this guide exists.
Connect Gmail and Google Calendar from Customize > Connectors in claude.ai or Claude Desktop, one connector at a time. Before you click Connect, write down three rules: the AI drafts and a human sends, it marks mail as read but never archives or deletes, and it never edits anything another person on your team owns. Keep every send action on "Needs approval." Test with a read-only question first.
Last verified against the live setup on 12 September 2026. Connector screens change; the rules do not.
Before I sold anything, I was an avionics technician on F-18s in the RCAF. Nobody on that flight line was ever handed the keys to the whole aircraft. You signed out the exact tools for the exact task from the tool crib, you did the job, and you signed every one of them back in. Not because the technicians were untrustworthy. Because the aircraft was worth more than anyone's confidence in their own memory. Your inbox is the aircraft. Scope is the tool crib.
What does connecting Claude to Gmail actually give it?
According to Anthropic's own Google Workspace connector documentation, once Gmail is connected Claude can search and read your email, manage labels and threads, list your saved drafts, and draft, send, reply to and forward messages. By default it asks for your approval before each of those sending actions. It cannot read attachment contents, only the fact that an attachment exists. Google Calendar goes further: view events and shared calendars, create, update and delete events, find mutual availability, manage attendee lists and respond to invitations.
Read that list again. "Send." "Delete events." "Respond to invitations."
One honest note, because you will find it if you look. At the time of writing, the connector page on claude.com still says Claude "cannot create, send, or modify emails," while the newer Help Center article says the opposite. The Help Center article is the current one. The lesson is not that the documentation is sloppy. It is that the capability list grew, and the permission screen you clicked through six months ago now covers more than it did. That is exactly why the rules have to live with you, not with the vendor.
This is also the difference between what we build for a client during an AI integration and what happens when someone connects everything on a Sunday night. The connection is identical. The rules around it are not.
Step 1: Write the rules before you click Connect
ATP AI, Sales and Marketing runs Claude against real business inboxes every day: quotes, invoices, client email, calendars with paid meetings on them. Here are the three standing rules that have kept it out of trouble. They are not clever. They are the rules a good executive assistant follows on their first week without being told.
- Drafts only. A human hits send. The AI writes the reply, saves it as a Gmail draft, and hands you the link. Every time. Nothing leaves the building without a human reading it. This one rule removes almost every way an AI can embarrass you in front of a customer.
- Mark as read only. When the AI triages the inbox, the only thing it is allowed to change is the read state. No archive, no delete, no spam-marking, unless you say so on that specific email in that specific conversation. An inbox you cannot find things in is worse than an inbox that is full.
- Never touch what another person owns. If someone on your team owns a thread, a deal, a calendar or a task, the AI recommends and the person acts. Automation owns its own notes and labels. Nothing else.
Write those three down somewhere the AI reads at the start of every session. Later in this post there is a prompt that turns them into a file you keep, and that file is the actual asset here. The connect button is free. The rules are the work.
Step 2: How to connect Claude to Gmail in claude.ai or Claude Desktop
The path is the same on the web app and the desktop app. If you have not installed the desktop app yet, the full install walkthrough covers every screen. The Gmail connector is available on Pro, Max, Team and Enterprise plans. Free accounts get chat, not this.
| # | Where you are | What to do |
|---|---|---|
| 1 | Claude sidebar | Click Customize, then Connectors. This is the page that lists everything Claude is connected to. |
| 2 | Connectors page | Click the + beside Connectors to open the directory. Find Gmail. |
| 3 | Gmail connector card | Read the description. It tells you what the connector can do. Then click Connect. |
| 4 | Google account picker | Pick the business account. Not your personal one. If the business account is not listed, sign in to it in the browser first. |
| 5 | Google permission screen | This is the screen everyone skips. It lists what Claude is asking for. Read every line. If it lists something you did not expect, stop here and read the next section before you continue. |
| 6 | Back in Claude | Gmail now shows as connected. Repeat steps 2 to 6 for Google Calendar. Each connector is separate, which means you can connect the calendar and leave the inbox alone if that is all you need. |
| 7 | Any chat | Click the + in the message box, hover Connectors, and toggle Gmail or Google Calendar on for that conversation. A connector that is connected but toggled off does nothing in that chat. |
Step 3: Set the approval level, and leave send on "Needs approval"
This is the part that makes rule one real instead of aspirational. Go back to Customize > Connectors, click the Gmail connector, and look at the permissions. On Team and Enterprise plans an Owner sets each connector to Always allow, Needs approval or Blocked, and decides whether members can change that. On individual plans, Claude asks before it sends, replies or forwards by default.
Keep it that way.
The temptation, about a week in, is to flip send to Always allow because the approvals feel slow. Do not. The approval prompt is the one place a human sees the message before a customer does. Ten seconds of friction per email is cheap. One wrong reply to your biggest account is not.
It is the combination of two things: a Google permission that covers read, send and permanent delete, plus a Claude setting where send needs no approval. Either one alone is survivable. Together, the AI can write and send an email on your behalf and then delete the evidence, with no human in the loop. Nobody sets that up on purpose. People set it up by clicking through two screens they did not read.
Step 4: Connect Google Calendar, then test it read-only first
Same path: Customize, Connectors, +, Google Calendar, Connect, business account, permission screen. Once both are connected, do not ask it to do anything yet. Ask it to look.
- "What is on my calendar tomorrow, and which of those meetings have an email thread I have not replied to?"
- "Find the last email from our accountant and tell me what they asked for."
- "Which unread emails from this week are from actual customers, not newsletters?"
If those three come back right, the connection works and you have not changed a single thing in either system. Only then move to writes, and start with the smallest possible one: "Draft a reply to that accountant email and save it as a draft. Do not send it." Open Gmail, check the draft is there, check nothing else moved. That is your test write. It is the same habit we teach every owner we set up: one record, verified together, before the real thing.
What we granted and what we declined at ATP
The official connectors ask for a fixed set of permissions. You accept the set or you do not connect. But the moment you go one step further, and run Claude through your own automation with your own Google token, you choose the permissions line by line. Google publishes the full list of Gmail permission scopes and Calendar permission scopes, with the exact wording each one shows on the consent screen. Here is what we chose for ATP's own inboxes, and why.
| Permission (Google's own wording) | Our call | Why |
|---|---|---|
| "View your email messages and settings" (read only) | GRANTED | The daily briefing and inbox triage need to read. No argument. |
| "Read, compose, and send emails" without permanent deletion (gmail.modify) | GRANTED | This is the one that lets the AI mark mail as read and apply labels. It explicitly excludes permanent delete, which is the whole point. The send ability it technically carries is fenced by rule one, in writing, in the AI's operating file. |
| "Manage drafts and send emails" (gmail.compose) | GRANTED | Drafts. The AI writes, saves the draft, hands over the link. A human sends. |
| "Send email on your behalf" as a standalone grant | DECLINED | There is no automation at ATP that needs to send without a human. So nothing gets a permission whose only job is sending. |
| "Read, compose, send, and permanently delete all your email from Gmail" (the full mailbox scope) | DECLINED | Permanent delete has no undo. No AI task we run has ever needed it. Refusing this one scope is most of the safety. |
| "View and edit events on all your calendars" | GRANTED | Creating and moving events is useful. Event edits are recoverable. |
| "See, edit, share, and permanently delete all the calendars you can access" | DECLINED | Sharing a calendar or deleting a whole calendar is a human decision. The narrower events-only scope does everything we actually need. |
Notice the pattern. Every "declined" is a permission whose worst case cannot be undone. Every "granted" is one where a mistake is a Ctrl+Z, not a phone call. That is the whole method. You do not need to understand OAuth to apply it. You need to ask one question of every line on the permission screen: if the AI gets this wrong, can I put it back?
Turn the three rules into a file the AI reads every time
Rules in your head do not survive the second week. Rules in a file the AI loads at the start of every conversation do. Paste this into Claude. It will ask you a few questions about how your inbox actually runs, then hand you back a finished file.
ROLE You are an operations assistant setting up standing rules for how an AI may work inside a business owner's Gmail and Google Calendar. CONTEXT The owner has just connected Claude to Gmail and Google Calendar. They want the AI to help with inbox triage, drafting replies, and calendar prep, and they want hard limits on what it may change. The default rules are: 1. Drafts only. The AI never sends. A named human sends. 2. Mark as read only. No archive, delete, or spam-marking without the owner naming the specific email in the current conversation. 3. Never edit anything another person on the team owns (their threads, their deals, their calendar, their tasks). Recommend, do not act. TASK Ask me, one at a time, up to five questions you need answered to make these rules fit my business. At minimum find out: who is allowed to send, which senders or labels the AI must never touch, what "urgent" means for us, which calendar it may create events on, and what it should do when it is unsure. Then write the finished rules file. FORMAT Return ONE complete markdown file in a single code block, named inbox-rules.md. It must contain: a one-paragraph purpose, the three default rules rewritten with my answers filled in, a short "never do" list, a short "always do" list, and a final line that says what the AI must do when a rule and my live instruction conflict (the live instruction wins, but it must say which rule it is overriding). Plain language. No jargon. QUESTIONS FIRST Do not write the file until I have answered your questions.
Paste that into Claude and it will ask you up to five questions before it writes anything. Answer them properly, in your own words. Then it hands you back a finished file built around your inbox and calendar, not around our example.
That file is the thing worth keeping. Save it as inbox-rules.md. The .md on the end just means markdown, which is a plain text file with a little structure, and it is the format AI tools read best. If that is new, we cover it in what an MD file actually is.
Now open a Claude Project and add that file under project instructions, or upload it as project knowledge. Every chat you start in that project begins already knowing it. You stop re-explaining yourself every morning, which is the entire point.
On a business Google account, there is one more step
If your email runs on Google Workspace (a company domain, not @gmail.com), the Google permission screen may refuse with an "access blocked" message before Claude ever gets in. That is not a Claude problem. Your Workspace admin has to trust the app once. Per Anthropic's documentation, the admin goes to admin.google.com, then Security > Access and data control > API controls > Manage third-party app access, clicks Add app, selects OAuth App Name, searches for Claude, and sets it to Trusted. Allow roughly fifteen minutes for it to take effect.
On Claude Team and Enterprise plans there is a matching switch on the Claude side: an Owner enables the Google Workspace connectors under Organization settings > Connectors before anyone on the team sees them. If the Gmail connector simply is not in your directory, this is why.
How to take the access back
You should know how to undo this before you need to. Two places, and you do both.
- In Claude: Customize > Connectors, click the Google connector, click Disconnect, confirm. Claude forgets the connection and will ask you to sign in again next time.
- In Google: go to your Google Account's third-party connections page, find Claude, click See details to read exactly what it was allowed, then Remove access and Confirm. Google's own instructions walk through the same screen.
Do the Google side even if you did the Claude side. Disconnecting in Claude tells Claude to stop. Removing access in Google tells Google to stop honouring the key. They are not the same thing, and only the second one is final.
When something goes wrong
Google says "This app is blocked"
On a Workspace account this is the admin trust step above, nearly every time. On a personal account it usually means the connector is asking for a permission Google has not verified for that app yet. Disconnect, wait, and reconnect. If it persists, it is on the vendor side, not yours.
Claude says "you have some access already" and will not reconnect
This shows up when you try to switch Google accounts on a connector that was already connected once. Claude sees the old grant and short-circuits. Fix it from the Google side: remove Claude's access on the Google connections page for the old account, then disconnect in Claude, then connect fresh and pick the right account.
The Google sign-in hangs after you click Continue
The permission was granted but the hand-back to Claude never landed, so the connector shows as connected and every request fails. Disconnect in Claude, remove access in Google, and run the connection again in a normal browser window rather than a private one.
Every request fails with "The caller does not have permission" or "Not connected"
The connection exists but the grant behind it is stale or partial. Same cure: remove on both sides, reconnect once. Do not reconnect five times in a row. Each attempt can leave another half-grant behind.
The Gmail connector is not in the directory at all
Plan or admin. It is available on Pro, Max, Team and Enterprise. On Team and Enterprise an Owner has to enable it under Organization settings first.
It archived or deleted something
Then rule two was not in a file it reads, or send and edit were set to Always allow. Put the rules in the file from this post. Set the connector back to Needs approval. Then check Gmail's Trash and All Mail, because archive is recoverable and, on the official connector, permanent delete is not something it does.
Frequently asked questions
Is it safe to connect Claude to Gmail?
It is as safe as the rules you put around it. The connection itself uses Google's standard sign-in and can be revoked from your Google Account at any time. The risk is not the connection. It is a send permission with no approval step, and an AI with no written rules. Fix those two and the rest is normal software.
Can Claude send emails from my Gmail?
Yes, with the current connector, and by default it asks for your approval before each send, reply or forward. Our recommendation is to leave that approval in place permanently and run on drafts. The AI writes, you send.
Does Claude read all my emails when I connect it?
No. It reads what it needs to answer the question you asked, when you ask it, and nothing runs in the background. It does not scan the whole archive on connection. Attachment contents are not readable at all through the connector, only the fact that an attachment exists.
Can I connect Claude to multiple Gmail accounts?
One Google account per connector at a time. Switching accounts is the single most common way people hit the "you have some access already" error. Remove the old grant on the Google side first, then reconnect with the new account.
Can Claude delete my calendar events?
The Google Calendar connector can create, update and delete events. Test read-only first, keep destructive actions on Needs approval, and put "never delete an event another person created" in your rules file. If you build your own automation, choose the events-only permission and decline the one that can delete whole calendars.
Why does Claude say the Gmail connector is not working after it was fine yesterday?
Almost always an expired or stale grant. Disconnect in Claude, remove access in Google, reconnect once. If it happened right after you changed your Google password or a Workspace admin changed a setting, that is the cause.
Is this different in Claude Desktop, Cowork or Claude Code?
The connection is made once, in Customize > Connectors, and every surface picks it up. Cowork and Claude Code inherit what you connected in the app. The rules file travels with your Claude Project, so it applies wherever you open that project.
The connection is free. The discipline is the product.
Every owner I have set this up for has said some version of the same thing afterwards: "I thought the hard part would be the technology." It never is. The connector takes ninety seconds. The permission screen takes thirty. The three rules take a page. And the two weeks of catching the AI doing slightly the wrong thing, correcting it in the file, and watching it stop, is the part that actually turns an inbox into a system.
Connect one thing. Read the permission screen. Write the rules down. Then give it two weeks on real email before you judge it.
Want it connected with the rules already written?
ATP AI, Sales and Marketing sets up Claude against your real inbox and calendar, on a call, with the permission choices and the standing rules built around how your business actually runs. Calgary based, working with owners across Canada, the United States, Mexico, Panama and the Philippines.
See how an AI integration works
Or book a call and bring the inbox you are worried about.
Written by Tyler Fikowski with AI assistance, and checked against ATP's live connector setup and the vendor documentation linked above on 12 September 2026.
Want more guides like this in your Google results?