Home › Blog › AI email assistant for Gmail

How to build an AI email assistant for Gmail that only reads

Akash Gupta Founder, AI Gurukul Last reviewed 23 October 2026 10 minute read

Almost everyone I meet has the same problem: too much email, and the one message that mattered hidden under forty that did not. It is also the perfect first AI Employee, because the job is clear, the result is easy to check, and if you keep it to reading, the worst it can do is sort a message into the wrong group.

I do not write code. I run my company with AI Employees in Claude Code, and each one is checked by a person every day. This is the build I use to teach the method. If you are new to the idea, start with what an AI Employee is.

What will this Gmail assistant do, and what will it never do?

Write this down before you connect anything. It is the assistant's job description.

• It reads the email you point it at, for example everything unread from the last 24 hours.
• It sorts each message into one of three groups:
  • Urgent: needs you today, such as a deadline, a payment problem, or a client who is waiting.
  • Needs a reply: someone asked you something, but it can wait a day.
  • For information: updates, receipts, newsletters. Nothing to do.
• It gives a reason for each one, in one short line, plus the sender and subject so you can find the email.
• It never sends, replies, forwards, drafts, labels, archives, marks as spam or deletes anything.
• You check it. You open the Urgent ones yourself before acting.

Can Claude read my Gmail?

Yes. Anthropic has an official Gmail connector. According to the Claude Help Center, checked on 23 October 2026, it can search and read email, draft email, send, reply to and forward email, read email details including attachment details, manage labels and threads, and list saved drafts. It cannot read what is inside attachments, only details like the file name. Source: Claude Help Center, Google Workspace connectors.

The same page says that by default Claude asks for your approval before each send, reply or forward. That is a good safety net, but it is not the same as read only. In our build, the sending tools are switched off completely, so there is nothing to approve by mistake.

Which Claude plan do I need?

• The connector itself: Anthropic says Google Workspace connectors, including Gmail, are available to all users. On Team and Enterprise plans, an owner must switch them on for the company first.
• Claude Code: if you want to run the assistant as a saved routine on one command, Claude Code needs a Pro, Max, Team or Enterprise plan. As of October 2026, Pro costs USD 20 a month, or USD 17 a month paid yearly.

The full version, with a saved routine and locked permissions, is in Claude Code. If Claude Code is new to you, read Claude Code for people who do not code first.

How do I connect Gmail to Claude?

These are the steps from Anthropic's help pages, checked on 23 October 2026. Menu names can move, so if something looks different, use the source links.

1. Open Claude in your browser or the desktop app and sign in.
2. Go to Customize, then Connectors. You can also click the + button at the bottom left of a chat, point at Connectors, and choose Manage connectors.
3. Find Gmail, read its description, and click Connect.
4. Google opens a sign in screen. Choose the Gmail account. Google will show what access Claude is asking for. Read it before you click Allow.
5. Back in Claude, the connector shows as connected. In a chat, turn it on from the Connectors menu under the + button.

If your email is a company Google Workspace account and the connection is blocked, your IT admin may need to mark Claude as a trusted app. The help page says this can take about 15 minutes to take effect. Source: Claude Help Center, using connectors.

Do not connect a work inbox to a personal Claude account without your company's permission. Some companies block this on purpose.

How do I make the Gmail assistant read only?

This is the most important step, and the one most guides skip. There are two places to do it.

In the Claude app

Anthropic's help page says you can switch off specific tools of a connector from the chat. Open the connector's tool list and switch off every tool that drafts, sends, replies, forwards, labels, moves or deletes. Keep only the tools that search and read. The help page does not spell out the exact clicks, so I cannot promise the menu looks the same for you. If you cannot find it, use the Claude Code method below, where the block is written down and checkable.

In Claude Code

Connectors you added in the Claude app show up in Claude Code by themselves when you are logged in with your Claude account. Type /mcp to see them and the list of Gmail tools. Source: Claude Code docs.

1. In your assistant's folder, type /permissions.
2. Add a deny rule for each Gmail tool that is not a read tool. In my own Claude Code on 23 October 2026, these had names like mcp__claude_ai_Gmail__send_message, mcp__claude_ai_Gmail__reply, mcp__claude_ai_Gmail__forward, mcp__claude_ai_Gmail__create_draft, mcp__claude_ai_Gmail__label_message and mcp__claude_ai_Gmail__trash_message. Your list may differ, so copy the names from your own /mcp screen.
3. Leave the read tools, the ones that search threads and get a message or thread, allowed.
4. Test it: ask Claude to "draft a reply to the newest email". It should be refused.

Why bother, when your instructions already say "never send"? Because Anthropic's docs are clear that permission rules are enforced by Claude Code itself, not by the AI, and a deny rule blocks the action in every mode. An instruction is a request. A deny rule is a lock. Source: Claude Code docs on permissions.

What instructions should I give my email assistant?

Save these as a skill, a saved routine, so you can run the whole job with one short command. In Claude Code, ask: "Make a skill called inbox-digest with these instructions." Then read what it wrote. The instructions I start learners with:

1. Search my Gmail for unread messages from the last 24 hours. Read each one. Do not change anything in my inbox.
2. Put each message in exactly one group: Urgent, Needs a reply, or For information.
3. Urgent means it needs me today: a deadline today or tomorrow, a payment problem, a client or family member waiting on me, or anything about security or my accounts.
4. For each message, write the sender, the subject, and one line on why you put it in that group.
5. Never guess. If you cannot tell, put it under Needs a reply and say you were not sure.
6. Never invent a deadline, an amount or a name that is not in the email.
7. End with a count: how many you read, and how many in each group.

Point 3 is yours to change. Your idea of urgent is not mine. Add the names of your key clients, your boss, or your bank, so the assistant learns what urgent means for you. The full six part method behind this is in how to build an AI agent with Claude, without coding.

What does the digest look like?

Here is the shape, made up for this example:

• Urgent. Priya, accounts team, "Invoice 214 rejected". Reason: says the payment will fail unless the GST number is fixed by tomorrow.
• Needs a reply. Rahul, "Slides for Thursday?". Reason: asks a direct question, meeting is in three days.
• For information. Airline, "Your e-ticket". Reason: a receipt, nothing to do.
• Count. Read 23. Urgent 1. Needs a reply 4. For information 18.

Then you open the Urgent email yourself. If the reason does not match what the email says, you fix the instructions.

What if I do not want to connect my real inbox?

Use a sample inbox. This is how I suggest most beginners start, and it teaches the same skill.

1. Make a new folder, for example "inbox-test".
2. Inside it, make a text file called sample-inbox.txt. Write or paste ten made up emails, each with a From line, a Subject line, a date and a short body. Mix them: two urgent, three that need a reply, five that are just updates.
3. Change step 1 of the instructions to: "Read the emails in sample-inbox.txt."
4. Run it. Because you wrote the emails, you know the right answer, and you can mark its work like a teacher.

Another option is a new, empty Gmail account with a few test emails in it.

How do I test it before I trust it?

1. Run it every morning for a week, and still glance through your inbox yourself.
2. Each day, note any message it put in the wrong group, and why.
3. Add one line to the instructions for each kind of mistake. For example: "Emails from my CA are always Urgent in the last week of the month."
4. Watch for the dangerous mistake: something urgent filed under For information. That is the one that costs you. If it happens, tighten point 3.

Is it safe to let AI read my email? Privacy cautions

• Your email holds other people's information. Clients, family and colleagues wrote to you, not to an AI. Keep the assistant to sorting, and do not paste its output into public places.
• Least access. Read only means read only. Do not switch sending back on "just this once".
• Sensitive mail. Bank, medical and legal email is still read when you connect. If that worries you, point the search at one label or one sender group instead of the whole inbox.
• Where data goes. Anthropic's help page notes that connected services handle data under their own terms. Read Claude's and Google's privacy terms for your plan.
• Disconnect any time. In Customize, then Connectors, choose Gmail and click Disconnect.

What are the honest limits of an AI email assistant?

• It can miss something urgent. AI sorting can be wrong, which is why you check the Urgent list and glance at the rest until you trust it.
• It cannot read attachments. The connector sees only details like the file name, so an urgent PDF with a vague subject may be missed.
• Big inboxes are slower. The help page notes that results can vary for large mailboxes, some advanced Gmail searches may not work, and Google's own limits apply.
• It runs when you ask. In this first version, you start it each morning. Running it on a timer is a later step.
• Things change. Connector tools and menu names change. This guide was checked on 23 October 2026.

For more jobs like this one, read how office workers use AI agents to get hours back, or the difference between this and a normal chat in AI agent or chatbot.

Akash Gupta, Founder, AI Gurukul

Akash studied at IIM Ahmedabad and built Peepoye, a Hindi YouTube channel. He does not write code. He runs his own company, Peepoye Network, with AI Employees he built himself with Claude, and teaches the same method in AI Gurukul. His story.

28.2 croreviews on his own channel

12.8 lakhsubscribers

1.4 crorehours watched