Skip to main content

​
Why start here

Before setting up any voice infrastructure, test your flow in a terminal. You get the full end-to-end dialog experience - same logic, same LLM calls, same tool execution - with nothing but Python. This is also the recommended eval loop:
  1. Build a flow
  2. Chat with it in the CLI
  3. Iterate on the prompt or flow structure
  4. Only then wire it into LiveKit / PipeCat / Unpod

​
Option 1: bundled CLI

If you already have a flow.json: 
superdialog chat kyc.json
That’s it. The CLI reads OPENAI_API_KEY / ANTHROPIC_API_KEY from your environment and runs the flow interactively. Override the model: 
superdialog chat kyc.json --llm anthropic/claude-haiku-4-5

​
Option 2: Python REPL loop

For more control - custom tool setup, custom model, traversal logging: 
import asyncio
from superdialog import create_dialog_flow, DialogMachine

async def main():
    flow = await create_dialog_flow(
        prompt="Confirm KYC. Ask for Aadhaar last 4. Confirm DOB.",
        llm="openai/gpt-5.1",
    )

    dialog_machine = DialogMachine(
        flow=flow,
        llm="anthropic/claude-haiku-4-5",
        traversal_dir="./traversal_history",  # saves a JSON per session on completion
    )

    while True:
        user = input("> ")
        if user.strip() in ("quit", "exit"):
            break
        reply = await dialog_machine.turn(user)
        print(reply.text)

asyncio.run(main())

​
Traversal history

Pass traversal_dir to save a timestamped JSON file for every completed session. Each file captures:
  • Every node visited (in order)
  • Every turn (user text + agent reply)
  • All collected slot values
dialog_machine = DialogMachine(
    flow=flow,
    llm="anthropic/claude-haiku-4-5",
    traversal_dir="./traversal_history",
)
Use these files to build eval corpora and inspect unexpected flow paths.

​
When to move to the next step

Once your flow behaves correctly in the CLI: