# Gmail MCP server: what it does, and when your agent needs its own inbox > Should your AI agent use a Gmail MCP server, or an email inbox of its own? ## The short answer A Gmail MCP server connects an AI assistant to your own Gmail account, so it can search your threads, draft replies and, depending on the server, send as you. That is the right tool for an assistant that works on your mail. An agent that does its own work, signing up for services, writing to suppliers or answering a support address, is better off with an inbox of its own, which it can reach over MCP in the same way. ## Three ways to give an assistant your Gmail There are three kinds of Gmail MCP server today, and they differ most in whether they can send. Every figure here is from the provider's own pages, read on 6 November 2026. **Google's own Gmail MCP server** is in Developer Preview, for members of the Google Workspace Developer Preview Program. It lists nine tools: search and read threads, create and list drafts, and list, add and remove labels. It does not send: the agent writes a draft and a person sends it from Gmail. It authenticates with OAuth and asks for two scopes, `gmail.readonly` and `gmail.compose`, through an OAuth client you create in Google Cloud. Its endpoint is `https://gmailmcp.googleapis.com/mcp/v1`. **Claude's Google Workspace connector** is Anthropic's own. It searches and reads mail, writes drafts, and can send, reply and forward, asking for your approval before each of those by default. Team and Enterprise owners turn connectors on for their organisation first. **Community servers** run on your machine with an OAuth client of your own and a token saved to disk. The most starred one on GitHub, archived in March 2026, could send with attachments, delete mail permanently, and create filters, all as you. ## What a Gmail MCP server is right for An assistant that helps **you** with **your** mail: summarising the morning's threads, finding the invoice from March, drafting a reply you will read before it goes. Your inbox is the point, the assistant acts on your behalf, and you are there to check it. Google's draft-only design and Claude's approval prompts both assume exactly this. ## Where it stops fitting The same server becomes awkward the moment the agent is doing a job of its own. - **It reads everything.** `gmail.readonly` is the whole mailbox: bank statements, password resets and every private thread. An agent that only needs supplier replies still holds the rest. - **It speaks as you.** Whatever it sends is from your address, so its mistakes arrive in your name, and the replies land in your inbox among your own mail. - **Every email is an instruction it might follow.** Anyone can write to your address, and a server that returns raw mail puts that text straight into the model's context. A Gmail MCP server has no idea which emails are trying to give the agent orders. - **It does not multiply.** Ten agents means ten Gmail accounts, or ten Workspace seats. A product that gives each customer an agent cannot hand each one somebody's Gmail. - **The limits are a person's.** Google limits a personal account that sends to more than 500 recipients in a day. And a public app that reads Gmail on other people's behalf uses restricted scopes, which means Google's verification, and a security assessment if the data reaches your servers. None of this is a flaw in the servers. They were built for a person's inbox, and they do that well. ## An inbox of its own, over MCP The other shape is an address that belongs to the agent. Its mail is only what was sent to it, it signs as itself, and making another one is an API call. With Pidgeon, the agent reaches it through a hosted MCP server, added in one line: ```sh claude mcp add --transport http pidgeon https://pidgeon.ai/api/mcp ``` The first call sends you to Pidgeon to sign in, where you choose what the client may do, read, send or full access, and which addresses it reaches. From then on the agent can create its own addresses, read and search them, reply in the thread, and wait for a verification code with one call that returns the code already pulled out of the email. What it adds for an agent's own work: - **Screened before it is read.** Every inbound message gets a verdict, and the text the tools return has hidden text removed. An address set to hold keeps suspicious mail from every key until a person releases it. - **A person where it matters.** An address can hold the agent's first email to someone new as a draft for a person to approve, and a reply on a thread with suspicious mail waits for a person on any address. - **A key per job.** A key reaches the addresses it was given and nothing else, and one made to read can never send. - **It says who sent it.** Everything the agent sends carries a signed statement that an agent wrote it and whom it acts for, which anyone can check. [The email MCP server page](https://pidgeon.ai/email-mcp-server) has the setup for Claude Code, Cursor, Codex, VS Code and Gemini CLI. ## Using both Most people who run agents end up with both, and the two do not overlap. The assistant that triages your mail keeps its Gmail connection. The agent that books, buys and follows up gets an address of its own, and when it needs you, it writes to you like anyone else would. Mail that arrives in Gmail for the agent can be forwarded to its address, so nothing has to move at once. ## Start To keep an assistant on your own mail, Google's [MCP configuration guide](https://developers.google.com/workspace/gmail/api/guides/configure-mcp-server) and Claude's Google Workspace connector are the official routes. To give an agent an address of its own, [sign up](https://pidgeon.ai/signup) and add the hosted server; [the MCP docs](https://pidgeon.ai/docs/mcp.md) cover every client and the full list of tools. ## Questions ### Is there an official Gmail MCP server? Yes, from Google, in Developer Preview for members of the Google Workspace Developer Preview Program. It searches and reads threads, writes drafts and manages labels, and does not send. ### Can a Gmail MCP server send email? It depends on the server. Google’s writes drafts only; Claude’s Google Workspace connector can send, reply and forward, asking for approval by default; community servers usually send as you. ### When should an agent use its own inbox instead of Gmail? When it does a job of its own: signing up for services, writing to suppliers or answering a support address. Then it should see only its own mail, send as itself, and be one of many. ### Can an agent use an inbox of its own over MCP? Yes. Pidgeon’s hosted MCP server adds in one line, and the agent can create addresses, read, reply and wait for verification codes in its own inbox.