Gmail MCP Server
Multi-account Gmail integration for AI assistants via the Model Context Protocol (MCP).
README
Gmail MCP Server
Multi-account Gmail integration for AI assistants via the Model Context Protocol (MCP).
Give your AI assistant secure access to all your Gmail accounts -- search, read, triage, draft, label, archive, trash-stage, and send emails with per-account routing, macOS Keychain token storage, and two-step confirmation flows for sensitive actions.
The Problem
Built-in Gmail integrations for AI assistants have significant limitations:
- Single account only -- no way to switch between personal, work, and business inboxes
- Insecure token storage -- most open-source servers store OAuth tokens as plaintext JSON files
- Overly broad scopes -- many request full
mail.google.comaccess (read, write, delete, send) - No send protection -- an AI can compose and send an email in a single step with no human checkpoint
What This Server Does Differently
| Concern | This Server | Typical MCP Gmail Server |
|---|---|---|
| Accounts | Unlimited, with aliases (personal, work) |
Single account |
| Token storage | macOS Keychain (encrypted at rest) | Plaintext JSON files |
| OAuth scopes | 4 scoped Gmail scopes plus scoped Docs/Drive (9 total, no mail.google.com) |
mail.google.com (full access) |
| Sending | Two-step: stage draft, then confirm with a 60-second code | Direct send, no confirmation |
| Trash/bulk changes | Two-step staged confirmation | Often direct mutation |
| Credentials | Bring your own Google Cloud OAuth client | Often bundled (security risk) |
| Shell injection | execFileSync with argument arrays |
execSync with string interpolation |
Features
- 35 tools for serious Gmail and Workspace workflows: search/read, batch read, thread reads, draft create/update/list, forwarding, labels, filters, archive, staged trash, staged bulk labeling, profile, attachments, two-step send, Google Docs/Drive reads, and scoped Drive writes
- Multi-account routing -- every tool accepts an optional
accountparameter (email or alias) - Active account -- set a default account that persists across restarts
- Rich draft bodies -- create/update drafts can include raw
html_bodywith a plain-text fallback for Gmail-safe formatted replies - Safe modify -- label/archive/trash operations require explicit message IDs or staged query preview/confirmation
- Attachment handling -- downloads to a sandboxed quarantine directory with SHA-256 integrity hashes and read-only permissions
- Minimal dependencies -- only the official Google APIs, MCP SDK, and Zod
Quick Start
Prerequisites
- Node.js 18+ (tested on 22)
- macOS (uses Keychain for token storage)
- A Google Cloud project with OAuth credentials
Install & Build
git clone https://github.com/arkaigrowth/gmail-mcp-server.git
cd gmail-mcp-server
npm install
npm run build
Smoke Test
echo '{"jsonrpc":"2.0","method":"tools/list","id":1}' | node dist/index.js
You should see a JSON response listing all 35 tools.
Add to Your AI Assistant
Claude Code (~/.claude/settings.json under mcpServers):
{
"gmail": {
"command": "node",
"args": ["/absolute/path/to/gmail-mcp-server/dist/index.js"]
}
}
Claude Desktop (claude_desktop_config.json):
{
"mcpServers": {
"gmail": {
"command": "node",
"args": ["/absolute/path/to/gmail-mcp-server/dist/index.js"]
}
}
}
Authenticate Your First Account
Once the MCP server is connected, tell your AI assistant:
"Authenticate my Gmail account user@gmail.com with alias personal"
This will:
- Open your browser to Google's consent screen
- Store the OAuth refresh token in your macOS Keychain
- Register the account for future use
You can then immediately start using it:
"Search my personal inbox for emails from GitHub in the last week"
Google Cloud Setup
You need a Google Cloud OAuth client (type: "Desktop app") with the Gmail API enabled. Each Google account needs its own OAuth client from a project where that account has access.
Step-by-step
- Go to the Google Cloud Console
- Create a new project (or use an existing one)
- Enable the Gmail API: APIs & Services > Library > search "Gmail API" > Enable
- Configure the OAuth consent screen: APIs & Services > OAuth consent screen
- User type: External (or Internal if using Google Workspace)
- Add the Gmail scopes:
gmail.readonly,gmail.compose,gmail.modify,gmail.settings.basic - Add the Workspace scopes used by the Docs/Drive tools:
documents.readonly,documents,drive.metadata.readonly,drive.readonly,drive.file - Add your email as a test user (while in "Testing" status)
- Create OAuth credentials: APIs & Services > Credentials > Create Credentials > OAuth client ID
- Application type: Desktop app
- Download the JSON file
- Save the JSON file as
~/.gmail-mcp/client_secret.json
Tip: If you have multiple Google accounts, you can specify a per-account path via the
client_secret_pathparameter when authenticating.
Credential File Lookup Order
client_secret_pathparameter passed togmail_authenticate- Path stored in the account's config (
accounts.json) GMAIL_MCP_CLIENT_SECRET_PATHenvironment variable~/.gmail-mcp/client_secret.json(default)
Tool Reference
Account Management
| Tool | Description |
|---|---|
gmail_list_accounts |
List all configured accounts and their auth status |
gmail_switch_account |
Set the active default account (persisted across restarts) |
gmail_authenticate |
OAuth flow -- opens browser, stores token in Keychain |
Email Operations
| Tool | Description |
|---|---|
gmail_search_messages |
Legacy search using Gmail query syntax (from:alice is:unread) |
gmail_search_emails |
Paginated rich search with messages, nextPageToken, and estimate |
gmail_read_message |
Read a specific message with full body content |
gmail_batch_read_messages |
Read explicit shortlisted message IDs, with partial failures |
gmail_read_thread |
Read a conversation thread, optionally capped with max_messages |
gmail_list_labels |
List all labels (system + custom) with message counts |
gmail_get_profile |
Get mailbox metadata (message/thread counts) |
Labels, Archive, Trash
| Tool | Description |
|---|---|
gmail_create_label |
Create a label, or return an existing label by name |
gmail_apply_labels_to_messages |
Add/remove labels on explicit message IDs only |
gmail_archive_messages |
Archive explicit message IDs by removing INBOX |
gmail_stage_trash_messages |
Stage explicit messages for Trash and return a confirmation code |
gmail_confirm_trash_messages |
Confirm staged Trash operation; account-bound, short expiry |
gmail_prepare_bulk_label_query |
Preview and stage capped query-based bulk labeling |
gmail_confirm_bulk_label_operation |
Confirm a staged bulk label operation |
Gmail filters
| Tool | Description |
|---|---|
gmail_create_filter |
Create a filter from criteria and label actions, with optional existing label-name resolution |
gmail_list_filters |
List filter IDs, criteria, and actions |
gmail_delete_filter |
Delete exactly one filter by ID |
Filter label names must already exist. The server returns available label names when a requested name cannot be resolved, and does not auto-create labels.
Compose & Send
| Tool | Description |
|---|---|
gmail_create_draft |
Create a draft email (does not send) |
gmail_list_drafts |
List draft metadata |
gmail_update_draft |
Update an existing draft; rejects drafts with attachments in v1 |
gmail_create_forward_draft |
Create draft forwards; quotes original content inline |
gmail_send_draft |
Stage a draft for sending -- returns a confirmation code |
gmail_confirm_send |
Confirm and send with the code (60-second expiry) |
Attachments
| Tool | Description |
|---|---|
gmail_download_attachment |
Download to quarantine directory with SHA-256 hash |
All email tools accept an optional
accountparameter (email or alias). If omitted, the active default account is used. If no default is set and only one account exists, it's auto-selected.
Draft bodies are plain text by default. For rich Gmail rendering, pass html_body with body as the plain-text fallback; the server emits multipart/alternative with text/plain first and text/html second. You can also set content_type: "text/html" to treat body as raw HTML when no separate fallback is available. Outbound attachments are not supported in draft create/update yet.
Security
Token Storage
OAuth tokens (refresh tokens, client credentials) are stored exclusively in the macOS Keychain using the security CLI. No plaintext token files are ever written to disk.
Keychain Service: gmail-mcp-server
Keychain Account: <your-email-address>
The ~/.gmail-mcp/accounts.json config file stores only metadata (email, alias, description, credential file path) -- never secrets.
OAuth Scopes
The consent flow requests one combined set of scopes (REQUIRED_AUTH_SCOPES in src/scopes.ts). It covers Gmail plus the Google Docs and Drive tools this server also exposes (reading shared docs and specs, and writing scoped Docs/Drive files), so the Workspace scopes are requested in the same flow.
Gmail scopes
| Scope | Grants |
|---|---|
gmail.readonly |
Read messages, threads, labels, profile |
gmail.compose |
Create and send drafts |
gmail.modify |
Label, archive, trash, and other safe mailbox state changes |
gmail.settings.basic |
Create, list, and delete Gmail filters |
Google Docs / Drive scopes
| Scope | Grants |
|---|---|
documents.readonly |
Read the text of Google Docs the account can open |
documents |
Create and edit Google Docs the account can access |
drive.metadata.readonly |
Read Drive file metadata (names, types), no file content |
drive.readonly |
Read and download the content of any Drive file the account can access, including Office files shared by teammates |
drive.file |
Create and manage the Drive files this server creates or opens |
The Gmail scopes are intentionally narrower than mail.google.com full access: the server does not request gmail.send; direct messages.send is blocked and all sends must go through draft staging plus confirmation. The Workspace scopes add Docs/Drive read and scoped-write access on top of that.
The scopes are defined as grouped constants in src/scopes.ts (REQUIRED_GMAIL_SCOPES, REQUIRED_WORKSPACE_READ_SCOPES, REQUIRED_WORKSPACE_WRITE_SCOPES). If you only need the Gmail tools, remove the Workspace groups from REQUIRED_AUTH_SCOPES and rebuild; the consent flow then requests only the Gmail scopes and the Docs/Drive tools stop working.
Tokens store the granted scope list in Keychain metadata. Tokens granted before a scope was added produce a clear reauth-required error for the tools that need it; rerun gmail_authenticate for that account.
Re-authentication required: Adding
gmail.settings.basicmeans each account must rerungmail_authenticateand complete the consent flow before filter tools work. Existing tokens keep working for old tools, but filter calls fail with insufficient scopes until re-authentication.
Two-Step Send Confirmation
Sending email requires two separate tool calls:
gmail_send_draft-- stages the draft and returns a 16-character cryptographic confirmation code (valid for 60 seconds)gmail_confirm_send-- the AI must present the code back, and the account must match
Additionally, messages.send is explicitly blocked on the Gmail client -- all sends must go through the draft confirmation flow.
Two-Step Trash And Bulk Label Confirmation
Trash and query-based bulk label operations use short-lived, account-bound confirmation codes. Explicit message ID label/archive operations are direct because their targets are already selected; query-based bulk labeling always previews before mutation.
OAuth Callback Security
- Callback server binds to
127.0.0.1only (never0.0.0.0) - Random
stateparameter validated to prevent CSRF - Callback server auto-terminates after 5 minutes by default (override via
GMAIL_MCP_AUTH_TIMEOUT_MS) - Dynamic port allocation (no fixed ports)
Attachment Quarantine
Downloaded attachments are:
- Saved to
~/.gmail-mcp/quarantine/<messageId>/with0o700directory permissions - Written with
0o400(read-only) file permissions - Filename sanitized to prevent path traversal
- Accompanied by a SHA-256 hash for integrity verification
Shell Injection Prevention
All calls to the macOS security CLI use execFileSync with argument arrays, not execSync with string interpolation. This prevents shell injection via crafted email addresses or token data.
Multi-Account Examples
Setup two accounts
"Authenticate user@gmail.com with alias personal"
"Authenticate work@company.com with alias work, client_secret_path ~/.gmail-mcp/work_client_secret.json"
Switch between them
"Switch to my work account"
"Search work for emails from the finance team this month"
"Switch to personal"
"Read my latest emails"
Or specify per-call
"Search my personal inbox for order confirmations"
"Search my work inbox for meeting invites tomorrow"
File Structure
~/.gmail-mcp/
accounts.json # Account metadata (no secrets)
defaults.json # Active account preference
client_secret.json # Default OAuth client credentials (from Google)
quarantine/ # Downloaded attachments (sandboxed)
<messageId>/
attachment.pdf # Read-only (0o400)
macOS Keychain
Service: gmail-mcp-server
Account: user@gmail.com # Token JSON, including granted OAuth scopes
Account: work@company.com # Token JSON, including granted OAuth scopes
Development
# Build
npm run build
# Run in development (with tsx)
npm run dev
# Run tests
npm test
# Smoke test (JSON-RPC)
echo '{"jsonrpc":"2.0","method":"tools/list","id":1}' | node dist/index.js
Project Structure
src/
index.ts # Entry point: MCP server + stdio transport
tools.ts # 35 tool registrations with Zod schemas
gmail-client.ts # Gmail API wrapper (search, read, draft, labels, filters, archive, trash, send, attachments)
accounts.ts # Account config management (load, resolve, persist)
keychain.ts # macOS Keychain integration via security CLI
oauth.ts # OAuth browser flow with localhost callback
scopes.ts # Gmail OAuth scope constants and metadata checks
types.ts # Shared TypeScript interfaces
skills/
gmail-multi/ # Repo-local workflow guidance for agents using this server
Platform Support
Currently macOS only due to the Keychain dependency. Contributions for Linux (secret-tool / libsecret) and Windows (wincred) credential storage are welcome.
License
MIT
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
Neon Database
MCP server for interacting with Neon Management API and databases
E2B
Using MCP to run code via e2b.
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.