MCP Server Boilerplate
A TypeScript boilerplate for building MCP servers with streamable HTTP and OAuth proxy support.
MCP Server Playground
A playground for Model Context Protocol (MCP) server built with TypeScript and Streamable HTTP transport with an OAuth Proxy for 3rd party authorization servers like Auth0.
Features
- MCP Server implementation: HTTP-Based Streamable transport using
@modelcontextprotocol/sdkwith HTTP transport, session management, and tool execution. - OAuth authentication/3rd party authorization: Implements an OAuth server for MCP clients to process 3rd party authorization servers like Auth0, providing Dynamic Application Registration for MCP server.
- Storage: Provide storage for MCP server to store data like OAuth sessions, tokens, etc.
- Session Management: Support stateful sessions by using replay of initial request.
- Tools:
aws-ecs: Investigate the ECS service, task and cloudwatch logs using AWS ECS, Cloudwatch Logs and Bedrockaws-s3: Get the list of S3 buckets and objectssystem-time: Get the current system time in various formats with timezone supportecho: Echo a message back with transformations, repetitionstreaming: Simulate real-time streaming data with live updatesproject: Find keywords in the current project directory
- Prompts:
echo
Why this project exists?
- The Model Context Protocol spec requires Dynamic Application Registration because it provides a standardized way for MCP clients to automatically register with new servers and obtain OAuth client IDs without user interaction. The main reason for this mechanism is because MCP clients can't know all possible services in advance and manual registration would create significant effort for users and it is not scalable. If do not support Dynamic Application Registration, then MCP clients need to provide OAuth client ID and secret to the server, which is not secure and not scalable.
- However, enabling Dynamic Application Registration (if supported) becomes a security risk because the endpoint is a public endpoint that anyone can create OAuth clients. It can easily be abused, such as by flooding with unwanted client registrations. Hence, Auth0 has disabled Dynamic Application Registration
- As a result, this project provides a way to enable Dynamic Application Registration for MCP server by using OAuth Proxy, but delegating authorization to 3rd party authorization server like Auth0, Github, Google, etc.
Endpoints
| Endpoint | Description |
|---|---|
| GET /ping | Ping the server |
| POST /mcp | MCP protocol request with authentication |
| DELETE /mcp | Session termination |
| GET /.well-known/oauth-authorization-server | OAuth authorization server metadata |
| GET /.well-known/oauth-protected-resource | OAuth protected resources metadata |
| POST /oauth/register | Register a new MCP client |
| GET /oauth/authorize | Handle authorization request |
| POST /oauth/token | Handle token request |
| POST /oauth/revoke | Handle token revocation |
| GET /oauth/stats | Get OAuth service statistics |
| GET /oauth/auth0-callback | Handle Auth0 callback |
Getting Started
Installation
-
Clone the repository:
git clone <your-repo> cd mcp-server-playground -
Install dependencies:
npm install -
Set up environment variables:
cp .env.example .env -
Set up the MCP server for local development
npm run dev:setup
Helm Chart
helm repo add chrisleekr https://chrisleekr.github.io/helm-charts/
helm repo update
helm install mcp-server-playground chrisleekr/mcp-server-playground
Set up the MCP server for Cursor
-
Create MCP configuration file for local build
Create a
.cursor/mcp.jsonfile in your project directory (for project-specific setup) or~/.cursor/mcp.jsonin your home directory (for global setup):{ "mcpServers": { "mcp-server-playground-cursor": { "type": "http", "url": "http://localhost:3000/mcp" } } }
Use npx @modelcontextprotocol/inspector to test the MCP server
-
Copy
mcp-config.example.jsontomcp-config.json -
Edit
mcp-config.jsonto point to the correct MCP server -
Run the inspector
npm run docker:run # Then run the inspector npx @modelcontextprotocol/inspector -y --config ./mcp-config.json --server mcp-server-playground-cursoror
npm run test:inspector
Setup Auth0 for authorization
-
Create a new application in Auth0
- Go to Auth0 Dashboard
- Click on "Applications"
- Click on "Create Application"
- Name: MCP Server Boilerplate
- Application Type: Regular Web Application
- Click on "Create"
-
Set up the application
- Click on "Settings"
- Set the following settings:
- Allowed Callback URLs:
http://localhost:3000/oauth/auth0-callback - Allowed Web Origins:
http://localhost:3000
- Allowed Callback URLs:
-
Create a new API
- Click on "APIs"
- Click on "Create API"
- Name: MCP Server Boilerplate
- Identifier:
urn:mcp-server-playground - JSON Web Token (JWT) Profile: Auth0
- JSON Web Token (JWT) Signature Algorithm: RS256
- Click on "Create"
How to make stateful session with multiple MCP Server instances?
When the MCP server is deployed as a cluster, it is not possible to make it stateful with multiple MCP Server instances because the transport is not shared between instances by design.
To make it truly stateful, I used Valkey to store the session id with the initial request.
When the request comes in to alternative MCP server instance, it will check if the session id is in the Valkey. If it is, it will replay the initial request and connect the transport to the server.
Inspired from https://github.com/modelcontextprotocol/modelcontextprotocol/discussions/102
The below diagram shows the flow of the stateful session management.
TODO
- Streaming is not working as expected. It returns the final result instead of streaming the data.
Screenshots
| Metadata Discovery | Client Registration | Preparing Authorization |
|---|---|---|
| Authorization with 3rd party server | Request Authorization and acquire authorization code | Token Request and Authentication Complete |
|---|---|---|
References
Related Servers
Model Context Protocol servers
A collection of reference implementations for the Model Context Protocol (MCP), showcasing servers built with TypeScript and Python SDKs.
SonarQube MCP Server
Integrates with SonarQube to provide AI assistants with access to code quality metrics, issues, and analysis results.
Linear Regression MCP
Train a Linear Regression model by uploading a CSV dataset file, demonstrating an end-to-end machine learning workflow.
Dart MCP
An example MCP server built with Dart and deployed on Cloudflare Workers.
Divvi MCP Server
Automatically integrate the Divvi referral SDK into JavaScript and TypeScript blockchain applications.
Infercnv-MCP
Infer Copy Number Variations (CNVs) from single-cell RNA-Seq data using a natural language interface.
Apple HIG
Provides instant access to Apple's Human Interface Guidelines, with content auto-updated periodically.
Moondream
A vision language model for image analysis, including captioning, VQA, and object detection.
My MCP Server
A remote MCP server deployable on Cloudflare Workers without authentication.
Windows Command Line MCP Server
Enables AI models to interact with the Windows command-line safely and efficiently.