Installation:npx @artymclabin/gmail-mcp auth - or just tell your Claude to install the MCP from this repo (https://github.com/ArtyMcLabin/Gmail-MCP-Server) and let it set up. Prefer manual steps? See .
Lean Gmail MCP server with auto authentication support
The original repository has been unmaintained since August 2025 - 7+ months with zero maintainer activity and 72+ unmerged pull requests. I use this MCP server daily as part of my Claude Code workflow and depend on it working correctly, so I picked it up.
Pull requests are welcome. If you've been sitting on fixes or features with nowhere to submit them, this is the place.
Philosophy
This fork is lean and pragmatic. It's a local stdio MCP server - you run it on your own machine, and your LLM client already has shell + filesystem access. So the threat model is "don't leak credentials to third parties, don't break the Gmail surface" - not "defend a hosted multi-tenant service". I keep dependencies minimal. I use this daily in my own Claude Code workflow - if I wouldn't run it or maintain it myself, it doesn't go in.
There's a downstream fork that took this in the maximalist direction. I'm not affiliated with its maintainer and I don't track its security or features - use it at your own risk: klodr/gmail-mcp. If that's the philosophy you want, go check it out. PRs welcome here as always.
What this fork adds
Fixed reply threading - auto-resolves In-Reply-To and References headers so email replies land in the correct thread instead of creating orphaned messages (upstream PR #91, still pending)
Send-as alias support - optional from parameter for multi-identity email management (send from any configured Gmail alias)
Reply-all tool - reply_all automatically fetches the original email, builds To/CC recipient lists (excluding yourself), and sets proper threading headers (PR #3 by @MaxGhenis)
Fixed list_filters - was returning empty array due to wrong response property name (PR #4 by @nicholas-anthony-ai)
Custom OAuth2 scoping - --scopes flag to request only the permissions you need, with automatic tool filtering (PR #6 by @tansanDOTeth)
CI/CD hardening - fixed shell injection vector in GitHub Actions workflow, added least-privilege permissions scope (PR #9 by @JF10R)
Security hardening - fixed path traversal in attachment download, restricted OAuth credential file permissions (PR #10 by @JF10R)
Dependency security - upgraded MCP SDK to v1.27.1 (3 CVE fixes), upgraded nodemailer (DoS + routing fix), moved dev-only packages out of production deps (PR #11 by @JF10R)
Thread-level tools - get_thread, list_inbox_threads, get_inbox_with_threads, modify_thread for efficient thread-based email operations in a single call
CC/BCC visibility - read_email now shows CC and BCC headers when present (PR #21 by @panghy)
Phishing report tools - report_phishing and batch_report_phishing for marking messages as spam via the Gmail API (PR #24 by @ShivamB25)
Draft lifecycle tools - send_draft, delete_draft, update_draft close the orphan-draft gap: send_draft atomically sends an existing draft and removes it from Drafts (no ghost copy); update_draft mutates a draft in place preserving its ID (no draft pile-up across iteration loops); delete_draft discards an abandoned draft (PR #30 by @thisisambros)
Tool annotations - MCP spec annotations (readOnlyHint, destructiveHint, idempotentHint) on all tools for safer LLM tool execution (PR #14 by @bryankthompson)
Download email tool - download_email saves emails to disk in json/eml/txt/html formats without consuming LLM context (PR #13 by @icanhasjonas)
Durable OAuth sessions - refresh_token is persisted across restarts, ending the hourly re-auth loop (PR #35 by @BrentBaccala)
Custom OAuth callback port - the auth listener derives port and path from your callback URL instead of hardcoding 3000 (PR #41 by @soapergem)
Safe permanent-delete gating - delete_email/batch_delete_emails require the opt-in gmail.full scope (which also satisfies all other mail scopes), so default auth stays least-privilege (PR #39 by @caioribeiroclw-pixel)
All features are production-tested in daily use.
A Model Context Protocol (MCP) server for Gmail integration in Claude Desktop with auto authentication support. This server enables AI assistants to manage Gmail through natural language interactions.
Features
Send emails with subject, content, attachments, and recipients
Full attachment support - send and receive file attachments
Download email attachments to local filesystem
Download full emails to files in json/eml/txt/html formats
Thread-level operations - get full threads, list inbox threads, batch-expand threads
Support for HTML emails and multipart messages with both HTML and plain text versions
Full support for international characters in subject lines and email content
Read email messages by ID with advanced MIME structure handling
Enhanced attachment display showing filenames, types, sizes, and download IDs
Search emails with various criteria (subject, sender, date range)
Comprehensive label management with ability to create, update, delete and list labels
List all available Gmail labels (system and user-defined)
List emails in inbox, sent, or custom labels
Mark emails as read/unread
Move emails to different labels/folders
Delete emails
Batch operations for efficiently processing multiple emails at once
Full integration with Gmail API
Simple OAuth2 authentication flow with auto browser launch
Support for both Desktop and Web application credentials
Global credential storage for convenience
Installation & Authentication
Installing from npm (recommended)
npx @artymclabin/gmail-mcp auth
Installing from source
git clone https://github.com/ArtyMcLabin/Gmail-MCP-Server.git
cd Gmail-MCP-Server
npm install
npm run build
Choose either "Desktop app" or "Web application" as application type
Give it a name and click "Create"
For Web application, add http://localhost:3000/oauth2callback to the authorized redirect URIs
Download the JSON file of your client's OAuth keys
Rename the key file to gcp-oauth.keys.json
Run Authentication:
You can authenticate in two ways:
a. Global Authentication (Recommended):
# First time: Place gcp-oauth.keys.json in your home directory's .gmail-mcp folder
mkdir -p ~/.gmail-mcp
mv gcp-oauth.keys.json ~/.gmail-mcp/
# Run authentication from anywhere
node dist/index.js auth
b. Local Authentication:
# Place gcp-oauth.keys.json in your current directory
# The file will be automatically copied to global config
node dist/index.js auth
The authentication process will:
Look for gcp-oauth.keys.json in the current directory or ~/.gmail-mcp/
If found in current directory, copy it to ~/.gmail-mcp/
Open your default browser for Google authentication
Save credentials as ~/.gmail-mcp/credentials.json
Note:
After successful authentication, credentials are stored globally in ~/.gmail-mcp/ and can be used from any directory
Both Desktop app and Web application credentials are supported
For Web application credentials, make sure to add http://localhost:3000/oauth2callback to your authorized redirect URIs
Custom callback URL / port: By default the local OAuth server listens on port 3000 at /oauth2callback. If port 3000 is unavailable, or you need a different redirect URI, pass a full callback URL as an argument. The listener automatically binds to the port and path from that URL:
Set up your n8n container to expose a port for authentication
Configure a reverse proxy to forward traffic from your domain (e.g., gmail.gongrzhe.com) to this port
DNS Configuration:
Add an A record in your DNS settings to resolve your domain to your cloud server's IP address
Google Cloud Platform Setup:
In your Google Cloud Console, add your custom domain callback URL (e.g., https://gmail.gongrzhe.com/oauth2callback) to the authorized redirect URIs list
This approach allows authentication flows to work properly in environments where localhost isn't accessible, such as containerized applications or cloud servers.
OAuth Scopes
You can limit the server's Gmail access by specifying OAuth scopes during authentication. This controls which tools are available to the LLM, reducing the attack surface for sensitive operations.
Available Scopes
Scope
Description
gmail.readonly
Read-only access to emails (search, read, download attachments)
gmail.modify
Full read/write access to emails (superset of readonly - includes sending, modifying, deleting)
gmail.compose
Create drafts and send emails only
gmail.send
Send emails only
gmail.labels
Manage labels only
gmail.settings.basic
Manage filters and settings
Note: gmail.modify is a superset that includes all read capabilities. You don't need gmail.readonly if you have gmail.modify.
Authenticating with Specific Scopes
Use the --scopes flag to request only the permissions you need:
# Read-only access (recommended for safe browsing)
node dist/index.js auth --scopes=gmail.readonly
# Read-only with filter management
node dist/index.js auth --scopes=gmail.readonly,gmail.settings.basic
# Full access (default behavior)
node dist/index.js auth --scopes=gmail.modify,gmail.settings.basic
If no --scopes flag is provided, the server defaults to gmail.modify,gmail.settings.basic for full functionality.
Scope-to-Tool Mapping
The server automatically filters available tools based on your authorized scopes:
gmail.full is intentionally separate from the default scopes because it grants permanent-delete capability. Prefer modify_email / batch_modify_emails for archive, mark-read, label, or inbox cleanup flows; re-authenticate with --scopes=gmail.full,gmail.settings.basic only when the assistant should be able to permanently delete mail. gmail.full is a superset of the other mail scopes, so that combination keeps every read/send/modify/label tool available (settings scopes remain separate - Gmail's filter endpoints only accept gmail.settings.*).
Re-authenticating
To change your scopes, simply run the auth command again with different scopes. This will replace your existing credentials.
Claude Code CLI Configuration
To use this MCP server with Claude Code, add it to your MCP settings.
Read-Only Configuration (Recommended for Safe Browsing)
First, authenticate with read-only scope:
node dist/index.js auth --scopes=gmail.readonly
Then add to your Claude Code MCP settings (~/.claude/mcp_settings.json or project-level .mcp.json):
This enables all 23 tools including sending emails, managing labels, creating filters, reply-all, thread operations, phishing reports, and batch operations.
Running multiple instances (tool-name prefix)
Some MCP clients dedupe tool entries by their base name across servers, which makes it impossible to run two instances of this server side-by-side (e.g. one for a personal account and one for a shared inbox) - only one instance's tools surface, even though both servers report as connected.
The server accepts an optional --tool-prefix=<value> CLI flag (or GMAIL_MCP_TOOL_PREFIX env var) that is prepended to every tool name at registration. Default is empty (fully backward-compatible).
Example: register two instances with distinct prefixes in Claude Code:
# Personal account
claude mcp add gmail-personal -s user \
-e GMAIL_CREDENTIALS_PATH=$HOME/.gmail-mcp/credentials-personal.json \
-- node /absolute/path/to/Gmail-MCP-Server/dist/index.js --tool-prefix=personal_
# Shared inbox
claude mcp add gmail-info -s user \
-- node /absolute/path/to/Gmail-MCP-Server/dist/index.js --tool-prefix=info_
Tools then surface as mcp__gmail-personal__personal_search_emails, mcp__gmail-info__info_search_emails, etc. - distinct at every layer. The dispatch handler strips the prefix and looks up the resolved name via getToolByName; only strips when the result is a known tool (so a prefix value that overlaps a real tool name cannot cause silent mis-dispatch).
The auth subcommand runs before the server starts and is unaffected - invoke it without --tool-prefix.
Available Tools
The server provides the following tools that can be used through Claude Desktop:
1. Send Email (send_email)
Sends a new email immediately. Supports plain text, HTML, or multipart emails with optional file attachments and inline images.
Basic Email:
{
"to": ["[email protected]"],
"subject": "Meeting Tomorrow",
"body": "Hi,\n\nJust a reminder about our meeting tomorrow at 10 AM.\n\nBest regards",
"cc": ["[email protected]"],
"bcc": ["[email protected]"],
"mimeType": "text/plain"
}
{
"to": ["[email protected]"],
"subject": "Meeting Tomorrow",
"mimeType": "text/html",
"body": "<html><body><h1>Meeting Reminder</h1><p>Just a reminder about our <b>meeting tomorrow</b> at 10 AM.</p><p>Best regards</p></body></html>"
}
Multipart Email Example (HTML + Plain Text):
{
"to": ["[email protected]"],
"subject": "Meeting Tomorrow",
"mimeType": "multipart/alternative",
"body": "Hi,\n\nJust a reminder about our meeting tomorrow at 10 AM.\n\nBest regards",
"htmlBody": "<html><body><h1>Meeting Reminder</h1><p>Just a reminder about our <b>meeting tomorrow</b> at 10 AM.</p><p>Best regards</p></body></html>"
}
Inline Images (embedded in HTML):
Embed images inside the HTML body so they render in place, instead of arriving as separate attachments. Reference each image from htmlBody by its cid, and supply it as a file path or base64 content. Works the same way on draft_email, update_draft, and reply_all.
{
"to": ["[email protected]"],
"subject": "Quarterly Report",
"body": "Revenue is up. See the chart below.",
"htmlBody": "<p>Revenue is up:</p><img src=\"cid:chart1\">",
"inlineImages": [
{ "cid": "chart1", "path": "/path/to/chart.png" }
]
}
2. Draft Email (draft_email)
Creates a draft email without sending it. Also supports attachments.
{
"to": ["[email protected]"],
"subject": "Draft Report",
"body": "Here's the draft report for your review.",
"cc": ["[email protected]"],
"attachments": ["/path/to/draft_report.docx"]
}
3. Read Email (read_email)
Retrieves the content of a specific email by its ID. Now shows enhanced attachment information.
{
"messageId": "182ab45cd67ef"
}
Enhanced Response includes CC/BCC headers (when present) and attachment details:
Permanently deletes an email. This tool requires gmail.full (https://mail.google.com/); the default gmail.modify scope is not enough for Gmail's permanent-delete endpoint.
Permanently deletes multiple emails in efficient batches. This tool requires gmail.full (https://mail.google.com/); the default gmail.modify scope is not enough for Gmail's permanent-delete endpoint.
Replies to all recipients of an email. Automatically fetches the original email to build the recipient list and sets proper threading headers (In-Reply-To, References, threadId).
How it works:
Fetches the original email by messageId
Builds To from the original sender (From header)
Builds CC from original To + CC, excluding your own email
Sets threading headers so the reply lands in the correct thread
Sends via the existing send_email pipeline (supports attachments, HTML, multipart)
{
"messageId": "182ab45cd67ef",
"body": "Thanks for the update, everyone. I'll review and get back to you.",
"mimeType": "text/plain"
}
With HTML and attachments:
{
"messageId": "182ab45cd67ef",
"body": "Plain text fallback",
"htmlBody": "<p>Thanks for the update. See attached notes.</p>",
"mimeType": "multipart/alternative",
"attachments": ["/path/to/notes.pdf"]
}
Parameters:
messageId (required): ID of the email to reply to
body (required): Reply body (plain text, or fallback when using multipart)
htmlBody (optional): HTML version of the reply body
mimeType (optional): text/plain (default), text/html, or multipart/alternative
attachments (optional): Array of file paths to attach
21. Modify Thread (modify_thread)
Atomically modifies labels on an entire thread (all messages at once). Solves the problem where archiving only the latest message leaves older messages in the inbox.
Reports a message as phishing using the closest public Gmail API behavior by applying the SPAM label.
{
"messageId": "182ab45cd67ef"
}
Note: The Gmail API does not expose the full native "Report phishing" workflow. This tool applies the SPAM label as the closest available approximation.
23. Batch Report Phishing (batch_report_phishing)
Reports multiple messages as phishing in efficient batches.
Atomically sends an existing draft via users.drafts.send and removes it from the Drafts folder in the same operation - no orphan/ghost draft left behind. Use after a draft_email (or update_draft) once the content is confirmed.
{
"draftId": "r-1234567890123456789"
}
25. Update Draft (update_draft)
Replaces a draft's content in place via users.drafts.update, preserving the draft ID. Critical for iteration loops (draft → user requests changes → re-draft) so Drafts doesn't accumulate N copies. Reuses the same MIME builder as draft_email, so attachment and threading semantics match.
You can combine multiple operators: from:[email protected] after:2024/01/01 has:attachment
Advanced Features
Email Attachment Support
The server provides comprehensive attachment functionality:
Sending Attachments: Include file paths in the attachments array when sending or drafting emails
Attachment Detection: Automatically detects MIME types and file sizes
Download Capability: Download any email attachment to your local filesystem
Enhanced Display: View detailed attachment information including filenames, types, sizes, and download IDs
Multiple Formats: Support for all common file types (documents, images, archives, etc.)
RFC822 Compliance: Uses Nodemailer for proper MIME message formatting
Supported File Types: All standard file types including PDF, DOCX, XLSX, PPTX, images (PNG, JPG, GIF), archives (ZIP, RAR), and more.
Email Content Extraction
The server intelligently extracts email content from complex MIME structures:
Prioritizes plain text content when available
Falls back to HTML content if plain text is not available
Handles multi-part MIME messages with nested parts
Processes attachments information (filename, type, size, download ID)
Preserves original email headers (From, To, Subject, Date)
International Character Support
The server fully supports non-ASCII characters in email subjects and content, including:
Turkish, Chinese, Japanese, Korean, and other non-Latin alphabets
Special characters and symbols
Proper encoding ensures correct display in email clients
Comprehensive Label Management
The server provides a complete set of tools for managing Gmail labels:
Create Labels: Create new labels with customizable visibility settings
Update Labels: Rename labels or change their visibility settings
Delete Labels: Remove user-created labels (system labels are protected)
Find or Create: Get a label by name or automatically create it if not found
List All Labels: View all system and user labels with detailed information
Label Visibility Options: Control how labels appear in message and label lists
Label visibility settings include:
messageListVisibility: Controls whether the label appears in the message list (show or hide)
labelListVisibility: Controls how the label appears in the label list (labelShow, labelShowIfUnread, or labelHide)
These label management features enable sophisticated organization of emails directly through Claude, without needing to switch to the Gmail interface.
Batch Operations
The server includes efficient batch processing capabilities:
Process up to 50 emails at once (configurable batch size)
Automatic chunking of large email sets to avoid API limits
Detailed success/failure reporting for each operation
Graceful error handling with individual retries
Perfect for bulk inbox management and organization tasks
Security Notes
OAuth credentials are stored securely in your local environment (~/.gmail-mcp/)
The server uses offline access to maintain persistent authentication
Never share or commit your credentials to version control
Regularly review and revoke unused access in your Google Account settings
Credentials are stored globally but are only accessible by the current user
Attachment files are processed locally and never stored permanently by the server
Troubleshooting
OAuth Keys Not Found
Make sure gcp-oauth.keys.json is in either your current directory or ~/.gmail-mcp/
Check file permissions
Invalid Credentials Format
Ensure your OAuth keys file contains either web or installed credentials
For web applications, verify the redirect URI is correctly configured
Port Already in Use
By default authentication uses port 3000; either free it up before running authentication, or run on a different port by passing a custom callback URL (e.g. node dist/index.js auth http://localhost:8080/oauth2callback)
If freeing port 3000, you can find and stop the process using that port
A custom callback URL must match one of the authorized redirect URIs registered in the Google Cloud Console
Batch Operation Failures
If batch operations fail, they automatically retry individual items
Check the detailed error messages for specific failures
Consider reducing the batch size if you encounter rate limiting
Attachment Issues
File Not Found: Ensure attachment file paths are correct and accessible
Permission Errors: Check that the server has read access to attachment files
Size Limits: Gmail has a 25MB attachment size limit per email
Download Failures: Verify you have write permissions to the download directory
Contributing
Contributions are welcome! Please feel free to submit a Pull Request.
Branch workflow
This repo uses a two-branch model:
main - stable. Only receives changes promoted from experimental after they're confirmed working. PRs are never merged directly into main.
experimental - staging / active development. All PRs are retargeted here and merged into experimental first.
Lifecycle of a contribution:
PR opened (any base)
→ retargeted to `experimental`
→ security audit + review + CI
→ merged into `experimental`
→ soak / verify on experimental
→ `experimental` promoted to `main` (maintainer confirms)
Open your PR against experimental when possible. If you target main, a maintainer will retarget it to experimental before merge.
CI requires README updates - every push to main and every PR must include a README.md change (even a version bump or changelog entry). This ensures documentation stays current as the codebase evolves.
To bypass for commits that genuinely don't need a docs update (dependency bumps, CI config changes), include [skip-readme] or [no-readme] in your commit message or PR title.
Running evals
The evals package loads an mcp client that then runs the index.ts file, so there is no need to rebuild between tests. You can load environment variables by prefixing the npx command. Full documentation can be found here.