Concierge Docs - Build & Deploy MCP Servers

Install Concierge

Start by installing the concierge-sdk package. This gives you the Concierge CLI for scaffolding projects, free cloud hosting, and built-in debugging tools.

Terminal

pip install concierge-sdk

Scaffold Your Project

Create a new MCP server with one command. The --chatgpt flag adds widget support for building ChatGPT Apps.

Terminal

concierge init --chatgpt

This creates a new folder with everything you need:

Project Structure

my-app/
├── main.py           # Your MCP server with widgets
├── assets/           # Widget source files (JSX, CSS)
├── requirements.txt  # Dependencies
├── settings.json     # Deploy config
└── README.md

Terminal

cd my-app

Deploy to Production

Deploy your app to the cloud with a single command. You'll get a live HTTPS URL instantly.

Terminal

concierge deploy

You're Live!

Your app is now running at https://<project>.getconcierge.app/mcp

Want to see logs? Add the --logs flag:

Terminal

concierge deploy --logs

Test & Evaluate

Before connecting to ChatGPT, test your tools and widgets with Concierge. Run AI-powered evaluations to ensure everything works correctly.

Open Concierge

getconcierge.app

Connect your server

Enter your deployment URL or select from deployed servers

Test tools manually

Browse tools and resources, execute them with custom inputs

Run evaluation

Use the Evaluate tab to run AI-powered test scenarios

Connect to ChatGPT

Add your MCP server as a ChatGPT connector to use it in conversations.

Enable Developer Mode

Go to Settings → Apps & Connectors → Advanced settings and enable Developer mode

Create a Connector

Click Create, paste your URL https://<project>.getconcierge.app/mcp, add a name and description

Start Chatting

Open a new chat, click More menu, add your connector, and start using your tools!

Build Widgets

Widgets are interactive UI components that render inside ChatGPT. The @mcp.widget() decorator binds HTML to a tool, so when ChatGPT calls it, the widget renders automatically.

main.py — Widget Example

from concierge import Concierge

mcp = Concierge("my-app", stateless_http=True)

PIZZA_HTML = """<!DOCTYPE html>
<html><body style="font-family: system-ui; background: #1a1a2e; 
      color: white; text-align: center; padding: 32px;">
  <h1>Pizza Map</h1>
  <p>Finding the best pizza near you!</p>
</body></html>"""

@mcp.widget(
    uri="widget://pizza-map",
    html=PIZZA_HTML,
    title="Show Pizza Map",
)
async def pizza_map(topping: str) -> dict:
    """Show pizza spots for a given topping."""
    return {"topping": topping}

Widget Parameters

Parameter	Type	Description
`uri`	str required	Unique identifier for the widget
`html`	str	Inline HTML content for the widget
`entrypoint`	str	Path to HTML in assets/entrypoints/ (for JSX widgets)
`url`	str	External URL to embed in iframe
`title`	str	Display name shown in ChatGPT UI

Pro tip: Use entrypoint for React/JSX widgets. Concierge auto-builds them with Vite + Tailwind on deploy.

Congratulations!

You've built your first ChatGPT App. Now go build something amazing!

Join Discord · GitHub

Install Concierge

Start by installing the concierge-sdk package. This gives you the Concierge CLI for scaffolding projects and free cloud hosting.

Terminal

pip install concierge-sdk

Scaffold Your Project

Create a new MCP server with one command:

Terminal

concierge init

This creates a new folder with everything you need:

Project Structure

my-server/
├── main.py           # Your MCP server
├── requirements.txt  # Dependencies
├── settings.json     # Deploy config
└── README.md

Terminal

cd my-server

Deploy to Production

Deploy your MCP server to the cloud with a single command:

Terminal

concierge deploy

You're Live!

Your MCP server is now running at https://<project>.getconcierge.app/mcp

Test & Evaluate

Test your tools with Concierge before connecting to clients. Run AI-powered evaluations to validate behavior.

Open Concierge

getconcierge.app

Connect your server

Enter your deployment URL or select from deployed servers

Test tools manually

Browse and execute tools with custom inputs

Run evaluation

Use the Evaluate tab for AI-powered testing

Define Tools

Add tools to your MCP server using the @mcp.tool() decorator:

main.py

from concierge import Concierge
from pydantic import Field

mcp = Concierge("my-server", stateless_http=True)

@mcp.tool()
async def add(
    a: int = Field(description="First number"),
    b: int = Field(description="Second number")
) -> int:
    """Add two numbers together."""
    return a + b

@mcp.tool()
async def get_weather(city: str) -> dict:
    """Get weather for a city."""
    return {"city": city, "temp": "72°F", "condition": "Sunny"}

if __name__ == "__main__":
    import uvicorn
    app = mcp.streamable_http_app()
    uvicorn.run(app, host="0.0.0.0", port=8000)

Connect to Clients

Use your MCP server with any MCP-compatible client:

✓

Claude Desktop

Add to claude_desktop_config.json under mcpServers

✓

Cursor

Configure in Cursor settings → MCP Servers

✓

ChatGPT

Use concierge init --chatgpt for widget support

You're all set!

Your MCP server is ready. Deploy changes anytime with concierge deploy

Join Discord · GitHub

Connect to Your Server

Open the Concierge Dashboard and connect to your MCP server.

Add Server

Enter your server URL or select from your deployed servers

Open Evaluation

Click the "Evaluate" tab in the studio sidebar

Configure LLM

Choose your preferred LLM provider to power the evaluation scenarios.

Supported Providers

✓

OpenAI

GPT-4o, GPT-4o-mini, GPT-4-turbo

✓

Anthropic

Claude Sonnet 4, Claude Haiku 4

✓

Google

Gemini 2.5 Pro, Gemini 2.5 Flash

Run Evaluation

The evaluation runs automated test scenarios against your MCP tools.

What Gets Tested

🔧

Tool Discovery

Automatically discovers all tools, resources, and prompts exposed by your server

🤖

AI-Generated Scenarios

LLM generates realistic test prompts based on your tool descriptions

💬

Multi-Turn Conversations

Simulates real user conversations with tool invocations

📊

Performance Metrics

Measures response times, success rates, and tool coverage

Review Results

Get a comprehensive evaluation summary with actionable insights.

Evaluation Summary Includes

Conversation Transcript — Full ChatGPT-style view of all interactions
Tool Call History — Every tool invocation with arguments and results
Success/Failure Rates — Per-tool and overall success metrics
Response Times — Latency measurements for each tool call
Coverage Analysis — Which tools were tested vs. available
Per-Tool Scores — Individual ratings for each tool's performance

⚡

Advanced Features

Sampling Support

For tools that require server-side AI sampling (e.g., ctx.session.create_message()), Concierge automatically handles sampling requests using your configured LLM.

Elicitation Support

Tools that request additional user input via elicitation are fully supported. Dynamic forms are rendered for you to provide the required information.

Ready to Evaluate?

Connect to your MCP server and start testing with AI-powered evaluations.

Get Started →