Read, search, send, and manage labels in a Gmail account. Requires a Google account with an app password.
2 fields · 1 secret
Credentials are injected into an ephemeral sandbox at boot and destroyed with it.
Light — starts in seconds
Sessions run up to 60 minutes before the sandbox is reclaimed.
60+ AI models
Chat with 60+ AI models on the same workflow — switch to a different model mid-conversation and re-run the same prompt, or use Compare mode to put several side-by-side.
How models use it and what it is built for.
The Gmail MCP server exposes a Google mailbox to an AI model as a set of tools — search the inbox, read a thread, draft and send a message, and move mail between labels. It turns "find the invoice Stripe sent last week and forward it to accounting" into a sequence the model can actually execute.
It authenticates with a Google account plus an app password rather than a full OAuth consent screen, which is what makes it practical to run in a sandbox. The trade-off is that an app password grants access to the entire mailbox: there is no per-label or read-only scope to fall back on.
Run it hosted here and the server starts in an ephemeral sandbox with your credentials injected at boot. Nothing is installed locally and the sandbox is destroyed when the session ends.
Concrete operations exposed as tools.
In order. Each step assumes the previous one worked.
App passwords only exist on accounts with 2-Step Verification turned on. Enable it at myaccount.google.com → Security before anything else.
Go to myaccount.google.com/apppasswords, create a password for "Mail", and copy the 16-character value. Google shows it once. This is the value the server expects — your normal account password will not work.
The server needs your full email address and that app password. Both are injected as environment variables into the sandbox at boot and are never written to the page or the deploy log.
Ask the model to list your five most recent subject lines before you let it send anything. If search works, authentication is correct.
What you provide before the sandbox starts.
Gmail AddressRequiredYour full Google email address, e.g. you@gmail.com.
Gmail App PasswordRequiredSecret16-character app password generated at https://myaccount.google.com/apppasswords (regular account password will not work; 2FA must be enabled).
Paste any of these into Agent Studio once the server is connected.
Search my inbox for unread mail from the last 3 days and summarise what needs a reply.
Find the most recent invoice from Stripe and tell me the amount and due date.
Draft a polite decline to the meeting invite from Priya — do not send it yet.
Label every message from noreply@ as "Automated" and archive them.
Known constraints, stated plainly.
Deploy the server hosted, then watch which tools a model actually reaches for — with full JSON input and output on every call. Switch models mid-conversation to compare how each one uses the same server.
About the Gmail MCP server.
Treat it as you would any third-party mail client. The app password is injected into an ephemeral sandbox and discarded when the session ends, but while the session is live the model can read and send anything in that mailbox. Use a dedicated account for automation rather than your primary inbox, and revoke the app password at myaccount.google.com/apppasswords when you are done.
OAuth requires a verified Google Cloud project, a consent screen and a redirect URL you control — impractical for an ephemeral sandbox. An app password gets a working connection in about two minutes. The cost is that the credential is all-or-nothing.
Yes, if the model decides to. Send is a tool like any other. In Agent Studio every tool call is shown with its full JSON arguments before and after execution, so you can watch what it does — but there is no built-in confirmation gate on send.
Yes, provided your administrator allows app passwords and 2-Step Verification is enabled on the account. Many managed Workspace tenants disable app passwords by default.