Remote MCP Server Submission Guide

This is a complete guide for submitting your remote MCP server to Anthropic's MCP Directory for broader distribution and discoverability.

Prerequisites

Before server submission, you should have:

A working and fully-tested remote MCP server
OAuth 2.0 authentication implemented (if authentication required)
All tools with proper safety annotations
Production-ready deployment
A dedicated support channel (email or web)
Provisioned test account with sample data
Comprehensive documentation

New to remote MCP development? See Getting Started with Custom Connectors Using Remote MCP first. For technical best practices and protocol details, see MCP Protocol Documentation.

Note: This guide covers remote MCP servers (cloud-hosted, HTTPS). For local desktop extensions, see Local MCP Server Submission Guide.

1. Directory Overview

What are the benefits of directory inclusion?

Discoverability and trust:

Listed in official Anthropic MCP Directory accessible from Claude.ai
Accessible to Claude users across all platforms (web, desktop, mobile)
Professional visibility for your service

User experience:

One-click connection from directory
Integrated with Claude's connector interface
Standardized presentation across platforms
OAuth flow handled seamlessly

Support and credibility:

Anthropic review of quality, security, and compliance
Listed alongside other vetted connectors
Community visibility and feedback
Professional distribution channel

What Claude platforms support remote MCP servers?

All major Claude platforms:

Claude.ai (web) - Full support with OAuth
Claude Desktop - Full support with OAuth
Claude Code - Direct connection from user's machine (with OAuth support)
Claude API - Integration support
Claude Mobile apps - Connector support

2. Mandatory Requirements

All requirements in this section are mandatory for directory approval. Missing any of these will result in rejection or revision requests.

Are safety annotations required?

YES - Every tool MUST have accurate safety annotations.

Required on every tool:

readOnlyHint: true - For tools that only read data
destructiveHint: true - For tools that modify data or have side effects

See MCP Protocol - Tool Annotations for complete schema and implementation details.

Not optional. This is a hard requirement derived from the MCP Directory Policy.

How to decide which annotation:

Tool Behavior	Annotation	Examples
Only reads data	readOnlyHint: true, destructiveHint: false	search, get, list, fetch, read
Writes/modifies data	destructiveHint: true, readOnlyHint: false	create, update, delete, send
Creates temp files	destructiveHint: true	Even temporary writes count
Sends external requests	destructiveHint: true	Emails, notifications, webhooks
Caches internally only	readOnlyHint: true	Internal optimization OK

Additional recommended annotation:

title - Human-readable tool name for UI display (improves user experience)

Do I need to provide test accounts?

YES - If your server requires authentication.

What to provide:

Test account credentials (username/password or API keys)
Sample data in account (necessary for functional testing)
Setup instructions for test environment
Access limitations (if any)

Test accounts should have:

Access to all tools being reviewed
Representative sample data
Appropriate permissions for full functionality testing
Active status throughout review period and beyond

How to provide:

Include credentials in submission form (ideally shared via secure method, such as 1Password link)
Ensure accounts remain active during and after review for periodic post-admission reviews
Provide sufficient access for comprehensive testing

Is OAuth 2.0 required?

YES - If your server requires authentication.

OAuth implementation requirements:

Must use OAuth 2.0 authorization code flow
Certificates from recognized authorities
Allowlist local MCP client (e.g. Claude Code, MCP Inspector) callback URLs:
- http://localhost:6274/oauth/callback
- http://localhost:6274/oauth/callback/debug
Allowlist Claude callback URLs:
- https://claude.ai/api/mcp/auth_callback
- https://claude.com/api/mcp/auth_callback
Proper redirect URI configuration

Common OAuth issues to avoid:

Invalid redirect URI errors (ensure both callback URLs are allowlisted)
HEAD requests without tokens (handle gracefully after OAuth flow)

Implementation guidance: See OAuth 2.0 Authorization Framework for complete OAuth implementation details.

Are there firewall requirements?

YES - Must allowlist Claude's IP addresses for claude.ai compatibility

For servers behind firewalls, allowlist IP addresses from https://docs.claude.com/en/api/ip-addresses.

Required for: Claude.ai and Claude Desktop

Not required for: Claude Code (connects directly from user's machine)

Important: IP allowlisting alone is not recommended as a security measure. Use OAuth 2.0 for authentication whenever possible.

What documentation is required?

Comprehensive server documentation with specific sections.

Recommended sections:

Server Description - Clear explanation of what your server does
Features - Key capabilities and use cases
Setup Instructions - How users connect and configure
Authentication - OAuth setup and requirements (if applicable)
Usage Examples - Minimum 3 working examples with prompts (required)
Privacy Policy - Link to full privacy policy
Support - How users can get help or report issues

Example documentation structure:

# [Your Service Name] MCP Server

## Description
[Brief description of service integration and capabilities]

## Features
- Feature 1: [description and value]
- Feature 2: [description and value]  
- Feature 3: [description and value]

## Setup
1. Visit the [Anthropic MCP Directory](https://claude.com/connectors)
2. Find and connect to [Your Service]
3. Complete OAuth authentication
4. Configure any required settings

## Authentication
This server requires OAuth authentication. You'll need:
- Valid [Your Service] account
- [Any specific permissions or account types]

## Examples
[See minimum 3 examples section below]

## Privacy Policy
See our privacy policy: https://your-domain.com/privacy

## Support
- Email: [email protected]
- Documentation: https://your-domain.com/mcp-docs
- Issues: https://github.com/yourcompany/mcp-server/issues

How many usage examples are required?

MINIMUM of three working examples demonstrating core functionality.

What qualifies as a good example:

Shows realistic user prompt/request
Demonstrates actual server functionality
Includes expected output or behavior
Clear and understandable workflow
Covers different capabilities

Example format:

## Examples

### Example 1: Search for documents
**User prompt:** "Find recent project reports in my workspace"
**What happens:**
- Server searches your workspace
- Returns matching documents with metadata
- Provides quick access links

### Example 2: Create new content
**User prompt:** "Create a new task list for the marketing campaign"
**What happens:**
- Server creates new task list
- Adds initial structure based on context
- Returns link to newly created list

### Example 3: Update existing data
**User prompt:** "Update the project status to 'In Progress' and add today's milestone"
**What happens:**
- Server locates the project
- Updates status field
- Adds milestone with current date
- Confirms changes made

Requirements:

Minimum 3 examples (no maximum)
Cover different tools/capabilities
Show realistic user interactions
Demonstrate value proposition
Include in server documentation

What are the production readiness requirements?

Server must be in General Availability (GA) status.

Production-ready means:

Server is stable and reliable in production
Not marked as "beta," "alpha," or "development"
All features fully implemented and tested
Proper error handling and graceful failures
Scalable infrastructure and monitoring
Complete documentation and support channels

Cannot be included: Beta versions, development servers, or limited-access services.

What technical requirements must be met?

Must meet the core technical compliance standards.

Transport and Performance:

Must support Streamable HTTP transport (SSE support may be deprecated)
Fast response times with high availability
Graceful error handling with helpful messages
Token-efficient responses (max 25,000 tokens per tool result)

Security and Data:

HTTPS/TLS with valid certificates
CORS properly configured for browser clients
Support for all required Claude client origins
Collect only data necessary for functionality
No collection of extraneous conversation data
Privacy-compliant data practices

3. Submission Process

How do I submit my remote MCP server?

Follow this step-by-step submission process:

1. Pre-submission checklist:

Verify mandatory requirements:

[ ] All tools have readOnlyHint OR destructiveHint annotations
[ ] OAuth 2.0 implemented (if authentication required)
[ ] Server accessible via HTTPS
[ ] Claude IP addresses allowlisted (if behind firewall)
[ ] Comprehensive documentation published
[ ] Privacy policy published and accessible
[ ] Dedicated support channels (email or web)
[ ] Test account ready (if authentication required)
[ ] Server is production-ready (GA status)

Test your server:

[ ] Works correctly from Claude.ai
[ ] Works correctly from Claude Desktop
[ ] Works correctly from Claude Code (if no IP restrictions)
[ ] OAuth flow completes successfully
[ ] All tools function as documented
[ ] Error messages are helpful and user-friendly
[ ] Performance is acceptable under load

2. Complete submission form:

Submission form: MCP Directory Server Review Form

Required information: Server details, documentation links, test credentials, examples (minimum 3), and contact information. Form provides a complete list.

While we strive to review every submission as quickly as we can, due to the amount of interest we cannot promise that we will accept your submission or respond to it individually.

4. Common Issues

What are the most common reasons for revision requests?

These are the top issues based on submission data:

1. Missing tool annotations

Problem: Tools missing required safety annotations
Fix: Add readOnlyHint or destructiveHint to ALL tools
Impact: Immediate rejection, requires code changes
Prevention: Validate all tools before submission

2. OAuth implementation issues

Problem: OAuth flow fails or has configuration errors
Common causes:
- Missing callback URLs in OAuth provider
- Invalid redirect URI configuration
- Firewall misconfiguration
Fix: Test OAuth flow thoroughly with MCP Inspector, Claude Code or Claude.ai
Impact: Cannot complete functional testing, delays approval

3. Incomplete documentation

Problem: Missing examples, unclear setup instructions, or missing required sections
Fix: Provide minimum 3 detailed examples and complete all documentation sections
Impact: Revision request, delays approval
Prevention: Follow documentation template exactly

4. Production readiness concerns

Problem: Server marked as "beta" or shows instability issues
Fix: Ensure GA status and stable production deployment
Impact: Cannot be included until production-ready
Prevention: Complete internal testing and stability validation

5. Privacy policy/support channel issues

Problem: Missing or inaccessible URLs for privacy policy/support channels
Fix: Provide proper privacy policy and support channels (email or web)
Impact: Immediate rejection due to policy requirement
Prevention: Verify privacy policy/support channel accessibility before submission