MCP Playground 🛠️🌩️

A Streamlit-based playground that lets you chat with large language models and seamlessly plug in external Multi-Server Command Protocol (MCP) tools. Spin up multiple FastMCP servers (Weather & Currency) alongside a Streamlit client, all orchestrated with Docker Compose. The client is provider-agnostic (OpenAI • Amazon Bedrock • Anthropic • Google Gemini • Groq) thanks to LangChain + LangGraph. Built with uv for lightning-fast dependency installation.

📖 Learn More

Want a deep dive into how it all works? Check out the detailed walkthrough in this Medium article: https://medium.com/@elkhan.alizada/your-own-ai-agent-playground-build-it-with-streamlit-langgraph-and-docker-4caeb6fe0ac4

🖥️🔌 Main Interface – Connected View

🏗️ Architecture

✨ Key Features

📂 Project Layout

mcp-playground/
├─ docker-compose.yaml          # One-command orchestration
├─ client/                      # Streamlit UI
│  ├─ app.py                    # Main entry-point
│  ├─ config.py                 # Typed settings & defaults
│  ├─ servers_config.json       # MCP endpoint catalogue
│  ├─ ui_components/            # Streamlit widgets
│  └─ ...
└─ servers/
   ├─ server1/                  # Weather Service MCP
   │  └─ main.py
   └─ server2/                  # Currency Exchange MCP
      └─ main.py

🚀 Quick Start

1 · Prerequisites

• Docker ≥ 24 & Docker Compose
• At least one LLM provider key (e.g. OPENAI_API_KEY) or AWS creds for Bedrock.

2 · Clone & Run

git clone https://github.com/your-org/mcp-playground.git
cd mcp-playground
docker compose up --build

⚡ Fast Builds with uv: All services use uv for dependency management, providing significantly faster Docker builds compared to traditional pip. Dependencies are cached and only reinstalled when pyproject.toml changes.

⚙️ Configuration

Default Settings

All runtime settings are concentrated in client/config.py and environment variables.

Default Model IDs:

MODEL_OPTIONS = {
    'OpenAI': 'gpt-4o',
    'Antropic': 'claude-3-5-sonnet-20240620',
    'Google': 'gemini-2.0-flash-001',
    'Bedrock': 'us.anthropic.claude-sonnet-4-5-20250929-v1:0',
    'Groq': 'meta-llama/llama-4-scout-17b-16e-instruct'
}

🎯 Dynamic Model Override

You can override the default model for any provider directly in the UI:

1. Select your provider from the dropdown
2. See the default model displayed
3. Enter a custom model ID in the "🎯 Custom Model" field
4. Leave empty to use the default

Examples:

• OpenAI: Override gpt-4o with o1 for reasoning tasks or gpt-4o-mini for cost efficiency
• Anthropic: Use claude-opus-4.6 or claude-sonnet-4.6 for latest models
• Google: Try gemini-2.0-flash-exp or gemini-exp-1206 for experimental features
• Bedrock: Use us.anthropic.claude-opus-4-6-v1 for Opus 4.6 or us.anthropic.claude-sonnet-4-6 for Sonnet 4.6
• Groq: Try different Llama models like llama-3.3-70b-versatile

➕ Dynamic MCP Server Management

Add custom MCP servers without editing servers_config.json:

1. Click "➕ Add MCP Server" expander in the sidebar
2. Fill in the server details:
   • Server Name: Unique identifier (e.g., MyCustomServer)
   • Server URL: Full endpoint URL (e.g., http://localhost:8002/sse)
   • Transport Type: sse or http
   • Timeout: Connection timeout in seconds (default: 600)
   • SSE Read Timeout: SSE-specific timeout (default: 900)
   • Headers (optional): JSON object with custom headers

     {"Authorization": "Bearer token"}
     

3. Click "Add Server" to add multiple servers
4. Servers persist in your session until removed

Server Sources:

• 📄 From Config: Static servers from servers_config.json (cannot be removed via UI)
• 📝 User Added: Dynamic servers added through UI (can be removed with 🗑️ button)

MCP endpoints can also be statically defined in servers_config.json for persistent configuration.

💬 Using the Playground

1. Select Provider · Pick your LLM in the sidebar and paste the corresponding credentials.
2. Customize Model (Optional) · Override the default model ID with any model from your provider.
3. Add MCP Servers (Optional) · Add custom MCP servers through the UI or use the default servers.
4. Connect MCP Servers · Click connect; available tools appear in the MCP Tools list.
5. Chat · Type a question.
   • If connected, the React agent decides whether to call an MCP tool (e.g. get_current_weather).
   • Otherwise it falls back to plain LLM chat.
6. Inspect Tool Calls · Tool invocations are streamed back as YAML blocks with inputs & outputs.

Try: "What will the weather be in Baku tomorrow and how much is 100 USD in AZN?"

🛠️ Included MCP Servers

Weather Service :8000

mcp = FastMCP("Weather Service", host="0.0.0.0", port=8000)

@mcp.tool()
async def get_current_weather(location: str) -> dict: ...

@mcp.tool()
async def get_forecast(location: str, days: int = 3) -> dict: ...

Currency Exchange :8001

mcp = FastMCP("Currency Exchange", host="0.0.0.0", port=8001)

@mcp.tool()
async def get_currency_rates(date: str = None) -> dict: ...

@mcp.tool()
async def convert_currency(amount: float, from_currency: str, to_currency: str, date: str = None) -> dict: ...

🙏 Acknowledgements

• LangChain
• LangGraph
• FastMCP
• Streamlit