gmail-mcp

Gmail for your AI assistant — several accounts at once, on a server you own.

日本語版 · 简体中文

gmail-mcp connects Gmail to Claude and any other MCP client. It can search and read mail, send and reply-all with quoted history, forward, handle attachments and inline images, and manage drafts, labels, and threads — across several Google accounts at the same time.

It runs as a remote server on your own Cloudflare Worker, so the same connection answers from Claude Code on a laptop, claude.ai in a browser, and Claude on a phone. Each connection signs in to one Google account, and the Google refresh token stays in your Cloudflare account.

Two things push people here. The Gmail connectors built into Claude and Google read mail and write drafts, but cannot send, and hold one Google account per assistant account. Servers that can send are usually local processes — fine at a desk, invisible from a phone.

How it compares

google_workspace_mcp is the most complete project here. It covers all of Workspace rather than Gmail alone, and it appends your Gmail signature and pulls attachments straight from a URL, neither of which gmail-mcp does. shinzo-labs/gmail-mcp reaches vacation responders, delegates, and S/MIME through its 64 tools; those live under gmail.settings.*, a scope gmail-mcp never requests, so they stay beyond its reach whatever happens to a grant.

Two design differences decide most of the rest. Routing accounts by a call argument lets one grant touch every connected mailbox, while binding the mailbox to the connection means a wrong argument reaches nothing. And on reading, the local servers decode every part as UTF-8: ISO-2022-JP and Shift_JIS mail arrives garbled, and long messages that Gmail stores as attachment blobs come back with an empty body.

Related MCP server: Gmail MCP Server

Deploy it

About ten minutes. You need a Cloudflare account, bun, and a Google account. A domain on the Cloudflare account is optional — without one the Worker answers on workers.dev.

1 · Create a Google OAuth client

PROJECT="gmail-mcp-$(openssl rand -hex 3)"
gcloud auth login
gcloud projects create "$PROJECT" --name="gmail-mcp"
gcloud config set project "$PROJECT"
gcloud services enable gmail.googleapis.com

Google exposes no API for the next two steps, so they happen in the Cloud console:

• OAuth consent screen → External, then under Audience press Publish app. Left in Testing, Google expires every refresh token after 7 days and each connection dies with its token. Published, the app shows an unverified-app warning at sign-in and serves up to 100 accounts.

• Credentials → Create credentials → OAuth client ID → Web application, with https://<your-host>/callback as an authorized redirect URI. Keep the client ID and secret.

<your-host> is the domain you point at the Worker, or the workers.dev hostname it gets otherwise. Deploying first and coming back to fill this in works — the guide the Worker serves at / shows the exact value.

2 · Deploy the Worker

The button copies the repository into your GitHub account, creates the KV namespace and the Durable Object, and asks for the four secrets. It deploys to workers.dev; a custom domain is attached afterwards under Settings → Domains & Routes.

From a terminal instead:

git clone https://github.com/mkpoli/gmail-mcp && cd gmail-mcp
bun install
bun run setup

bun run setup asks which domain to answer on, creates or reuses the OAUTH_KV namespace, takes the client ID and secret, generates a cookie key, and deploys. Those first two answers land in wrangler.local.jsonc, which git ignores — wrangler.jsonc names no account's namespace and no one's domain, so a clone deploys anywhere. Re-running setup to rotate a single secret is safe.

3 · Connect a client

Leave the client ID and secret fields empty — MCP clients register themselves.

claude mcp add --transport http gmail-personal https://<your-host>/mcp
claude mcp add --transport http gmail-work     https://<your-host>/mcp/work

Run /mcp in Claude Code to sign each connection in to its Google account. In claude.ai it is Settings → Connectors → Add custom connector with the same URL. Any single-segment label works after /mcp/, which is how one deployment serves several mailboxes to clients that reject two servers sharing a URL.

Your deployment serves this guide at https://<your-host>/.

What it can do

whoami search_messages get_message get_thread get_attachment

send_message reply_all forward_message create_draft update_draft send_draft delete_draft list_drafts stage_attachment_begin stage_attachment_append stage_attachment_finish

list_labels create_label update_label delete_label modify_labels modify_thread_labels batch_modify_messages trash_message · untrash_message trash_thread · untrash_thread

Messages leave the way a mail client sends them: plain text with an HTML alternative, file attachments, and inline images referenced by cid:, nested as multipart/mixed › multipart/related › multipart/alternative. Subjects and display names use RFC 2047, filenames use RFC 2231, so Japanese, Chinese, and emoji survive the trip.

reply_all reads the original's Reply-To, From, To, and Cc, drops your own address and any address you send mail as, answers from the one the sender wrote to, carries the References chain, and quotes the original in whichever parts you send. forward_message reproduces the forwarded envelope and can re-attach the original's files.

create_draft with replyToMessageId writes the reply as a draft to edit before sending: it joins the original's thread, carries In-Reply-To and References, derives the reply-all recipients and the Re: subject, and quotes the original. update_draft changes only the fields it is given; recipients, text, files added by hand in any client, and the thread the draft answers are read back and kept. A file whose base64 will not fit through tool arguments is staged instead: stage_attachment_begin returns an upload URL that takes the raw bytes in one curl -T, stage_attachment_append takes base64 in chunks, and every attachments field accepts the resulting stagingId.

Reading is bounded on purpose: message and thread bodies have character budgets, a whole response has a byte ceiling, and an attachment is returned inline only while it stays small enough to read. A long mailing-list thread, or a large file, comes back trimmed with a note saying so rather than filling the assistant's context.

How it works

Two OAuth flows meet in one Worker. The MCP client authenticates to the Worker; the Worker authenticates to Google on your behalf. Neither side holds the other's credentials.

sequenceDiagram
    autonumber
    participant C as MCP client<br/>(Claude Code · claude.ai)
    participant W as Worker<br/>(OAuthProvider + McpAgent)
    participant G as Google<br/>(OAuth + Gmail API)

    C->>W: POST /register (dynamic client registration)
    C->>W: GET /authorize (PKCE challenge)
    W->>C: approval dialog
    C->>G: consent screen — pick the account
    G->>W: GET /callback?code=…
    W->>W: allowlist check on the verified email
    W->>G: exchange code → access + refresh token
    W->>C: MCP access token (Google tokens sealed inside the grant)
    C->>W: POST /mcp — tools/call
    W->>G: Gmail REST (token refreshed as needed)
    G->>W: message / thread / label data
    W->>C: tool result

Built with

• TypeScript on Cloudflare Workers — Durable Objects hold one MCP session each, KV holds the OAuth grants

• Hono — routing for the OAuth endpoints, the Google callback, and the setup page at /

• @cloudflare/workers-oauth-provider — the OAuth 2.1 server that MCP clients register against

• agents — McpAgent, the MCP transport over Durable Objects

• @modelcontextprotocol/sdk with Zod — tool definitions and argument validation

• Bun, Biome, Wrangler — install, test, lint, deploy

Gmail itself is called over plain fetch against the REST API. The official googleapis SDK assumes Node and carries far more than a Worker should ship, so message building, MIME parsing, and token refresh live in src/gmail.ts and src/utils.ts instead.

Endpoints

Who can sign in

ALLOWED_EMAILS decides, checked against the address Google reports as verified — after consent, before any grant exists.

Each grant reaches only the mailbox that authenticated it, so widening this list never widens access to mailboxes already connected. Setting * lets strangers use your deployment, and your Google client's quota, for their own mail.

Limits

Two ceilings keep a shared deployment from being drained, both set in wrangler.jsonc:

On the Workers Free plan a further ceiling applies: 50 outbound requests per invocation. A wide read spends one per message, so search_messages and list_drafts want maxResults at 45 or below there; above it the surplus comes back as per-message errors rather than results. The paid plan allows 1000.

Raise either and redeploy. Cloudflare's rate limiter reads its ceiling from the binding at build time, so the simple.limit on each is the only place that changes it. A single-user deployment can leave both alone — normal assistant use sits far below them.

Security

Self-hosting moves the trust question rather than removing it, so here is where everything sits.

• Your tokens stay yours. Refresh tokens are encrypted inside their OAuth grant in your KV namespace. A session's Durable Object holds the hour-lived access token, and the MCP agent framework keeps a copy of the grant there for as long as the object lives, refresh token included. Both stores are your own Cloudflare account, encrypted at rest. Mail is never stored — it passes through.

• One session, one mailbox. The MCP session is bound to the account that opened it, so a grant for one mailbox cannot act on another through a borrowed session id.

• Scope minimalism. gmail.modify covers reading, sending, labels, and trash. It excludes permanent deletion and all of gmail.settings.*, keeping auto-forwarding rules and filter exfiltration — the classic mailbox backdoors — outside what any stolen grant could do. Two read-only scopes are requested alongside it, userinfo.email and userinfo.profile: they are how the allowlist and the session binding know which account signed in, and they reach no mail.

• Headers cannot be smuggled. Every outgoing header value is rejected if it contains CR, LF or NUL, so no argument can break out of its own field to append one — a Bcc inside a subject line, say. Media types are validated, and quoted history is HTML-escaped. What this does not do is police the arguments themselves: bcc is a real parameter, so a model acting on an instruction hidden in a message body could still fill it in, and your client's approval prompt remains the check on that.

• Access can be withdrawn. Narrowing ALLOWED_EMAILS stops new sign-ins. A single account's access is revoked at myaccount.google.com/connections. Rotating the Google client secret invalidates every grant at once.

The Worker decrypts mail in memory while serving a request, as any hosted relay must. If that is unacceptable for a particular mailbox, run a local MCP server for that one.

How it was tested

253 unit tests cover message construction (MIME nesting, RFC 2047 folding, RFC 2231 filenames, CR/LF rejection, base64 wrapping), body extraction across charsets, reply and forward composition, the Google token flows, the sign-in allowlist, the CSRF and state-binding checks that guard the browser side of sign-in, and the tools themselves against a stand-in Gmail — session ownership, recipient composition, attachment selection, and what a partly-failed read returns.

Beyond that, every tool has run against real Gmail accounts, with a separate account checking what arrived:

Development

bun run dev     # wrangler dev on :8788
bun run check   # biome + tsc
bun test        # 253 unit tests
bun run assets  # regenerate the light and dark diagrams
bun run deploy

Questions and bugs

Open an issue.

License

Copyright © 2026 mkpoli. Released under the MIT License.

src/workers-oauth-utils.ts is derived from the remote-mcp-github-oauth demo in cloudflare/ai, Copyright © 2025 Cloudflare, Inc., used under the MIT License. See THIRD-PARTY.md.