Quickstart: Model Routing and Authentication

Route your first LLM call through Attune's model registry and confirm the authentication strategy is configured correctly.

from attune.models import get_auth_strategy, EmpathyLLMExecutor

strategy = get_auth_strategy()
print(strategy.to_dict())

Expected output (values reflect your local config):

{
  "subscription_tier": "pro",
  "default_mode": "auto",
  "prefer_subscription": true,
  "cost_optimization": true,
  "setup_completed": true,
  ...
}

If setup_completed is false, complete step 1 before continuing.

Prerequisites

The project is cloned and installed locally
An Anthropic API key or Claude subscription

Step 1: Set up your authentication strategy

Run the interactive setup to configure your subscription tier and cost preferences:

from attune.models import configure_auth_interactive

strategy = configure_auth_interactive(module_lines=1000)
strategy.save()

The CLI equivalent is:

python -m attune.models auth setup

Step 2: Run a task through the executor

Create an EmpathyLLMExecutor and run a task. The executor automatically selects the best model for the task type:

from attune.models import EmpathyLLMExecutor

executor = EmpathyLLMExecutor(provider="anthropic")
response = executor.run(
    task_type="user_query",
    prompt="Summarize the purpose of adaptive model routing.",
)

print(response.model_id)       # which model was selected
print(response.cost_estimate)  # estimated cost in USD
print(response.content)        # the response text

Step 3: Inspect routing statistics

Check how the AdaptiveModelRouter scored models for your workflow:

from attune.models import AdaptiveModelRouter, get_telemetry_store

router = AdaptiveModelRouter(telemetry=get_telemetry_store())
stats = router.get_routing_stats(workflow="my_workflow", days=7)
print(stats)

What you just did

Loaded your AuthStrategy and confirmed it is configured
Ran a prompt through EmpathyLLMExecutor, which selected a model from the registry based on task type
Queried AdaptiveModelRouter for routing statistics drawn from historical telemetry