Skip to main content

Option A: Scaffold a New Project

1

Create the project

2

Run it

Your MCP server is now running with staged workflows out of the box.

Option B: Convert an Existing MCP Server

Already have a FastMCP server? Change one import:
All your existing @app.tool() decorators keep working. You’ve now got access to stages, transitions, state, and provider modes.

Build a Shopping Workflow

Let’s build an e-commerce flow with three stages: browse, cart, and checkout.

What Happens at Runtime

Notice the LLM only sees 1-2 tools per stage instead of all 3 at once. At scale (50+ tools), this dramatically improves reliability and reduces costs.

Add a Provider Mode

Want to reduce context even further? Switch to code mode:the agent writes Python instead of making individual tool calls:
Now the LLM sees a single execute_code tool and writes scripts like:

Deploy

Your server gets a public URL and appears in the Concierge Platform with logs, analytics, and evaluation.

Next Steps

Transitions

Deep dive into stages and transitions

State

Learn about persistent session state

Backends

Plain, Search, Plan, and Code execution modes

Platform

Inspector, logs, evaluation, analytics