Don't Install an MCP Blindly

Set and test Gmail permissions on Claude to control what AI can do with your email.

Introduction

30 Second Summary

Every time you connect an AI tool to your email, you hand it some level of access to your personal messages. The permissions screen flies by in seconds, and it is easy to click Allow without a second thought.

In this project, you will connect Gmail to Claude, review every permission it requests, and set each one to a level you deliberately chose. Then you will test those settings live to prove Claude actually respects them.

What You'll Build

Picture opening a chat with Claude, asking it to search your inbox, and knowing with certainty what it can and cannot do because you tested every permission yourself.

By the end of this project, you'll have:

A Gmail connection on Claude where reads, drafts, and other actions each have a permission level you chose.
Real proof that your permission settings work by watching Claude obey, ask for approval, and refuse actions during a live test.
A personal rule of thumb for deciding when an AI action is safe to allow, when it needs a gate, and when it should be fully blocked.
Secret Mission: Try every trick you can think of to get Claude to bypass a blocked permission.

Are there any prerequisites?

No prerequisites needed. All you need is a web browser, and this project will walk you through setting up your Claude and Gmail accounts from scratch.

Before We Start

Before you connect anything, let's get clear on what you're about to build and why it matters.

Get Your Accounts Ready

In this project we use Gmail as our example connector, but the permission skills you learn here apply to any MCP. If you already use a different connector, you will still follow along with Gmail and apply the same thinking to yours later.

In this step, you will confirm that you can log in to both Claude and Gmail. If you do not have an account for either one yet, you will create one.

In this step, get ready to:

Confirm access to your Claude account at claude.ai.
Confirm access to your Gmail account at gmail.com.

Log in to Claude

Open a new tab in your browser.
Go to claude.ai.
Log in with your existing account.

✔️ I see my Claude conversations

You are all set. Any Claude plan works for this project: Free, Pro, Max, Team, or Enterprise.

ⓧ I don't have a Claude account yet

On the claude.ai page, follow the prompts to create a new account.
You can sign up with an email address or a Google account.
Once you have created your account, confirm you can see the main conversations page.

Which Claude plan should I pick?

Any plan works for this project, including the Free plan. You do not need to pay for anything to complete every step.

Log in to Gmail

Open another tab in your browser.
Go to gmail.com.
Log in with your existing account.

✔️ I see my inbox

Your Gmail is ready.

ⓧ I don't have a Gmail account yet

On the Gmail sign-in page, follow the prompts to create a new Google account.
Once you have created your account, confirm you can see your inbox.

Verify both accounts are ready

You should now have two browser tabs open. One shows your Claude conversations at claude.ai. The other shows your Gmail inbox at gmail.com.

Having trouble logging in?

If you cannot log in to Claude, try resetting your password from the sign-in page. For Gmail, use Google's account recovery flow at accounts.google.com.

Help me troubleshoot my account login.

Both accounts are ready. Next up, you will connect Gmail to Claude and choose exactly what it can and cannot do with your email.

Connect Gmail and Set Your Permissions

You are logged in to both Claude and Gmail. Now it is time to connect them so Claude can work with your email.

Most people click Connect and never look at what they just allowed. In this step, you will see exactly what Claude can access and set every permission to a level you deliberately chose.

In this step, get ready to:

Connect your Gmail account to Claude through the connector directory.
Review Google's OAuth permissions screen and understand what it actually allows.
Set every Gmail permission to a level you chose instead of leaving the defaults.

Connect your Gmail account to Claude

Claude uses connectors to talk to external services like Gmail. You will add the Gmail connector from Claude's connector directory.

✔️ I haven't connected Gmail yet

In Claude, click the + button in the lower left corner of your chat.
Hover over Connectors, then select Manage connectors.
Click the + button next to Connectors to open the connector directory.
Find the Gmail connector under the Web category and click it.
Review what it says it can do: search and read emails, draft emails, and manage labels.
Click Connect.

Google will ask you to sign in and show an OAuth permissions screen. This screen lists what Claude is requesting access to.

Read the screen before you click anything.
Notice that it mentions email sending permissions.

What about the email sending permission?

The OAuth screen mentions email sending, but Claude only reads your emails and creates drafts. The send function is not enabled.

Every email must be sent manually from your Gmail account. Claude cannot send anything on its own.

Click Allow to complete the connection.

You will be redirected back to Claude. The Gmail connector now appears in your connector list.

Connection not working?

If you are on a Google Workspace account managed by your organization, your admin may need to approve Claude first. Ask your IT admin to go to Security > Access and data control > API controls > Manage third-party app access and set Claude as Trusted.

If you are on a Team or Enterprise Claude plan, an Owner or Primary Owner must enable connectors at the organization level before you can authenticate.

Help me troubleshoot my Gmail connection to Claude.

ⓧ I already connected Gmail

You are already set up. Skip ahead to the next section where you will review and configure your permission levels.

If you connected Gmail a while ago and never checked the permissions, this is your chance to fix that.

Configure your permission levels

Gmail is connected, but every permission is still at its default setting. Defaults are not decisions. You are going to review each permission and set it to the level that matches your comfort.

Click Customize in the sidebar.
Select Connectors.
Click on the Gmail connector.

You will see its Tool permissions, organized by category. Each permission can be set to one of three levels.

What are the three permission levels?

Each permission level controls how much freedom Claude has to perform a given action. Read through all three before you start setting them.

Always allow

Claude performs the action without asking you first. Use this for safe, read-only actions that cannot change anything in your account: searching your inbox, reading an email, or viewing email metadata.

Nothing set to this level can cause damage, so getting out of the way lets Claude work faster.

Needs approval

Claude asks you before every action. You see exactly what it wants to do and say yes or no. Use this for actions that create, change, or organize things: creating a draft, managing labels, or organizing threads.

One click of approval is worth it when something in your account is about to change.

Blocked

Claude cannot perform this action at all, no matter what you ask. Use this for anything you never want an AI doing with your email.

If an action is off the table for you, put it fully off the table.

Go through each permission in the Tool permissions panel.
For each permission, select the level that matches your comfort: Always allow, Needs approval, or Blocked.

A recommended starting point

If you are not sure where to start, try this:

Set read-only actions (searching your inbox, reading emails, viewing metadata) and creating drafts to Always allow. Drafts are safe because they sit in your Drafts folder until you manually send them.
Set anything that creates or deletes something that is not a draft (managing labels, organizing threads) to Needs approval.
Set anything you never want Claude doing to Blocked.

Before you move on, think about the permissions you just set. If you asked Claude to delete an email right now, what do you think it would do based on your choices?

Scroll through every permission in the Tool permissions panel one more time.
Confirm that each one shows the level you deliberately chose, not a default you ignored.

Not sure if a setting saved?

Navigate away from the Gmail connector settings, then click back into Customize > Connectors > Gmail. Your permission levels should still show the values you selected.

Help me check whether my Gmail permission settings saved correctly in Claude.

Your Gmail connection is locked down with permissions you chose on purpose. Next up, you will put those settings to the test and see Claude obey them in real time.

Test Your Permission Settings

Your Gmail permissions are configured, but you have not seen them in action yet. A setting you never tested is a setting you are only guessing about.

In this step, you will run two live tests inside a Claude chat to prove your permission levels actually work. You will watch Claude read your email freely and then ask for your approval before creating a draft.

In this step, get ready to:

Test a read-only Gmail action to confirm your Always Allow setting works.
Test a draft action to confirm your Needs Approval setting works.

Open a new chat and test a read-only action

Claude's connectors work inside conversations. To test your permissions, you need to start a chat and ask Claude to interact with your Gmail.

Start a new chat in Claude.
Ask Claude to search your inbox by typing this message:

Search my inbox for the most recent email I received today.

Press Enter to send the message.

What should I see?

What happens next depends on the permission level you set for read-only actions in the previous step.

If you set read-only actions to Always Allow, Claude accesses your inbox and returns results without asking for permission first.
If you set read-only actions to Needs Approval, Claude shows an approval prompt before accessing your inbox. Review the request, then approve it to continue.

Claude's response should include details about a recent email in your inbox. This confirms your read permission level is working as you configured it.

Claude is not accessing your Gmail?

Make sure your Gmail connector is still connected. Navigate to Customize, then Connectors, and confirm Gmail shows as connected.
Try refreshing the Claude page and starting a new chat.
If your browser blocks third-party cookies (common in Safari), Claude may have trouble maintaining the Google connection. Try using Chrome with default cookie settings.

Claude is not using my Gmail connector when I ask about my emails. How do I fix this?

Test a draft action

You have confirmed reads work. Now test an action that creates something new in your Gmail account.

In the same chat, ask Claude to draft a reply to the email it found by typing this message:

Draft a reply to that email saying thank you.

Press Enter to send the message.

What should I see?

If you set draft actions to Needs Approval, Claude pauses and describes what it wants to do before creating the draft. Read the description carefully, then approve the action.
If you set draft actions to Always Allow, Claude creates the draft without pausing.

Claude only creates drafts. It cannot send emails. You would need to open Gmail, find the draft, and send it yourself.

If you approved the action, Claude confirms that a draft has been created and saved. You can find it in your Gmail Drafts folder.

You have now tested both permission levels live. Claude accessed your inbox according to your read settings and paused for approval before creating a draft. Every permission you configured is working exactly as you chose it.

Secret mission

Try to Trick Claude

Your permissions are set and tested, but can Claude be talked into breaking its own rules? Put your blocked permissions under pressure with creative prompt engineering and find out whether the safety rails really hold.

Clean Up Your Resources

This project runs entirely in your browser with free accounts, so there are no ongoing costs. Decide whether to keep your Gmail connector active, pause it, or disconnect it entirely.

Resources you used:

Gmail connector on Claude (connected via OAuth, with permissions configured).
Test chat conversations in Claude (contain retrieved email data stored on Anthropic's servers).

Keep everything running

No action needed. Choose this if you plan to keep using Claude with Gmail.

Your Gmail connector stays connected with the permission levels you configured.
Your test chat conversations remain in Claude and still hold any email data that was retrieved during testing.
Your Claude and Gmail accounts stay active.

Pause - I'll come back to this later

Turn off the Gmail connector for new conversations without fully disconnecting it. Your permissions and connection stay saved.

In any Claude chat, click the + button in the lower left corner.
Hover over Connectors and toggle the Gmail connector off.
When you want to use it again, toggle it back on from the same menu. Your permission settings will still be there.

Delete - I don't want to use this again

Disconnect the Gmail connector and remove any retrieved email data from Anthropic's servers.

Delete your test chat conversations in Claude. This removes any email data that was retrieved during those chats from Anthropic's servers.
Navigate to Customize > Connectors.
Find the Google connector and click it to view its details.
Click Disconnect. This revokes Claude's OAuth access to your Gmail account.
Your Claude and Gmail accounts remain active. This project did not create either account, so there is nothing else to remove.

Nice Work!

You just took full control of what an AI can and cannot do with your email. Most people never check their permissions. You tested yours live and proved they work.

Here's what you learned:

Connected Gmail to Claude through an OAuth flow, read the permissions screen instead of skipping it, and set every permission to a level you deliberately chose.
Tested your permission settings live by watching Claude obey Always Allow, pause for Needs Approval, and refuse Blocked actions.
Built a personal rule of thumb that transfers to every connector you add: reads can lean toward Always Allow, anything that changes, sends, or deletes belongs on Needs Approval or Blocked.
Tested Claude's safety rails by trying to trick it into bypassing a blocked permission with creative phrasing, and confirmed the guardrails hold.

Ready to quiz yourself?