This guide covers local development where you configure LLMs via config files and have full control over the execution environment.
Want a faster setup? Check out the Platform Quickstart - no LLM config files needed and built-in observability! Make sure you’ve completed the Installation steps before proceeding. 1. Create LLM Config File
Create a llm-config.override.jsonc file to configure your LLM models. This file will override the default configuration.
llm-config.override.jsonc
// Your custom LLM configuration
{
"planner": {
"provider": "openai",
"model": "gpt-5-nano"
},
"orchestrator": {
"provider": "openai",
"model": "gpt-5-nano"
},
"cortex": {
"provider": "openai",
"model": "gpt-5",
"fallback": {
"provider": "openai",
"model": "gpt-5"
}
},
"executor": {
"provider": "openai",
"model": "gpt-5-nano"
},
"utils": {
"hopper": {
// Needs at least a 256k context window
"provider": "openai",
"model": "gpt-5-nano"
},
"outputter": {
"provider": "openai",
"model": "gpt-5-nano"
}
}
}
.env file in your project root with necessary API keys:
# LLM API Keys (only include the ones you need)
OPENAI_API_KEY=your_key_here
XAI_API_KEY=your_key_here
OPEN_ROUTER_API_KEY=your_key_here
GOOGLE_API_KEY=your_key_here
# Optional: For local LLMs or custom OpenAI-compatible endpoints
# OPENAI_BASE_URL=http://localhost:1234/v1
Never commit your .env file to version control. Add it to your .gitignore.
Creating Your First Automation
Let’s write a simple script that opens a calculator app and performs a basic calculation.
import asyncio
from minitap.mobile_use.sdk import Agent
from minitap.mobile_use.sdk.types import AgentProfile
from minitap.mobile_use.sdk.builders import Builders
async def main():
# Create an agent profile
default_profile = AgentProfile(
name="default",
from_file="llm-config.override.jsonc"
)
# Configure the agent
agent_config = Builders.AgentConfig.with_default_profile(default_profile).build()
agent = Agent(config=agent_config)
try:
# Initialize the agent (connect to the first available device)
agent.init()
# Define a simple task goal
result = await agent.run_task(
goal="Open the calculator app, calculate 123 * 456, and tell me the result",
name="calculator_demo"
)
# Print the result
print(f"Result: {result}")
except Exception as e:
print(f"Error: {e}")
finally:
# Always clean up when finished
agent.clean()
if __name__ == "__main__":
asyncio.run(main())
Run the script
python calculator_demo.py
Initialize the Agent
The agent connects to your device and starts required servers.
Execute the Task
The agent interprets your goal, navigates the UI, and performs the calculation.
Clean Up
Resources are properly released.
Getting Structured Output
Mobile-use SDK can return structured data using Pydantic models:
import asyncio
from pydantic import BaseModel, Field
from minitap.mobile_use.sdk import Agent
from minitap.mobile_use.sdk.types import AgentProfile
from minitap.mobile_use.sdk.builders import Builders
# Define a model for structured output
class CalculationResult(BaseModel):
expression: str = Field(..., description="The mathematical expression calculated")
result: float = Field(..., description="The result of the calculation")
app_used: str = Field(..., description="The name of the calculator app used")
async def main():
# Create an agent
default_profile = AgentProfile(
name="default",
from_file="llm-config.override.jsonc"
)
agent_config = Builders.AgentConfig.with_default_profile(default_profile).build()
agent = Agent(config=agent_config)
try:
agent.init()
# Request structured output using Pydantic model
result = await agent.run_task(
goal="Open the calculator app, calculate 123 * 456, and tell me the result",
output=CalculationResult,
name="structured_calculator"
)
if result:
print(f"Expression: {result.expression}")
print(f"Result: {result.result}")
print(f"App used: {result.app_used}")
finally:
agent.clean()
if __name__ == "__main__":
asyncio.run(main())
Using Pydantic models ensures type-safe, validated output from your automation tasks.
Understanding the Code
Agent Profile
default_profile = AgentProfile(
name="default",
from_file="llm-config.override.jsonc"
)
AgentProfile defines which LLM models power different components of the agent.
Agent Configuration
agent_config = Builders.AgentConfig.with_default_profile(default_profile).build()
Builders.AgentConfig provides a fluent API to configure your agent.
Running Tasks
result = await agent.run_task(
goal="Your instruction here",
output=YourPydanticModel, # Optional
name="task_name" # Optional
)
Next Steps
