MCP Chat UI

TypeScript/React chat UI for interacting with agents and MCP tools via WebSocket. This project is not just a Chat UI — it is a boilerplate for building Model Context Protocol (MCP) servers and easily testing their capabilities by chatting with them. The architecture enables rapid development, extension, and interactive experimentation with MCP agents and tools.

https://github.com/user-attachments/assets/ce39f244-9d77-4d29-b6e2-1119063653b8

Agent & LLM Integration

Agents are run using the OpenAI Agents SDK for seamless integration with MCP. The current implementation uses OpenRouter to access various LLMs for agent responses. OpenRouter provides access to multiple AI models including OpenAI, Anthropic, Google, and many others through a unified API. This flexibility allows you to experiment with different large language models as needed.

Features

• Chat with AI agents and tools in real time
• Extensible tool system (register new tools by asking Copilot)
• During chat, users can see all tool calls with their arguments and responses for debugging purposes
• Chart rendering and data visualization
• Coding AI Canvas implementation, allowing users to create designs and interactive visualizations by providing a natural language description (like Gemini Canvas). 'f.e. a user could say "visualize inserting sorting algorithm" and the system would generate and show a interactive visual representation of the algorithm.'
• Google authentication integration
• Hot reload and fast development with Vite

Getting Started

Prerequisites

• Node.js (v18+ recommended)
• npm

Environment Setup

1. Copy .env.template to .env:

   cp .env.template .env
   

2. Set your OPENROUTER_API_KEY in the .env file to enable OpenRouter LLM features. You can get your API key from OpenRouter.
3. (Optional) Set GOOGLE_CLIENT_ID in .env to enable Google authentication for the chat UI.

Install Dependencies

# In project root
npm install
cd client
npm install

# or

npm run install_all

Run Development Server

# In project root
npm run dev
open http://localhost:5173/

Start Production Build

# In project root
npm run start
open http://localhost:3000/

Inspect MCP Server

# In project root
npm run inspect

Registering New MCP Tools

You can add new server-side tools by simply asking Copilot:

register new tool with name my_tool_name, input parameters param1: string, param2: number, output fields result: string

Copilot will generate the file and code for you. See .github/copilot-instructions.md for full details.

Registering External MCP Servers

You can add new external MCP servers by editing the mcp.json file in the project root. List your server configuration in this file to make them available for selection in the chat UI.

Example mcp.json

{
  "mcpServers": {
    "playwright": {
      "description": "Playwright MCP",
      "command": "npx",
      "args": ["@playwright/mcp@latest", "--headless", "--isolated"]
    },
    "memory": {
      "description": "Server Memory MCP",
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-memory"]
    }
  }
}

Project Structure

• client/ — React UI, organized by feature (chat, chart, common, ui)
• src/server/ — WebSocket server, agent/tool orchestration
• src/tools/ — MCP tools (auto-registered)
• src/agents/ — Agent logic
• .github/copilot-instructions.md — AI agent instructions and conventions

About

This project demonstrates a flexible, extensible chat UI for agent/tool workflows, and serves as a boilerplate for building and interactively testing MCP servers and tools. Developer productivity is powered by Copilot and clear conventions for rapid extension.