appworld-sdk 0.1.1

Client SDK for App World — RL & Eval environments for Android apps

App World SDK

Lightweight Python client for the App World platform — RL training and evaluation on real Android apps, with a standard Gymnasium interface.

Installation

pip install appworld-sdk

Requirements: Python >= 3.11

Quick Start

from appworld_sdk import AppWorldClient

client = AppWorldClient(api_key="YOUR_API_KEY", base_url="https://api.appworld.dev")

# Browse available tasks
tasks = client.list_tasks()

# Allocate device resources
env_info = client.create_environment(task_ids=["FennoteCreateNote"])
env_id = env_info["id"]

# Get a Gymnasium-compatible environment
env = client.make_env(env_id=env_id, env_type="eval")

# Run a task
obs, info = env.reset(task_name="FennoteCreateNote")
done = False
while not done:
    action = {"action_type": 0, "x": 0.5, "y": 0.5, "text": "", "direction": 0}
    obs, reward, terminated, truncated, info = env.step(action)
    done = terminated or truncated

env.close()

Environment Types

Type	Class	Actions	Use Case
"rl"	RemoteAppWorldEnv	5 discrete (tap, swipe, etc.)	RL training
"eval"	RemoteAppWorldEvalEnv	7 (incl. text input)	Agent evaluation

Client-Side Eval with Custom Agents

Implement the BaseAgent protocol to run evaluations where LLM calls stay on your side — your model API key never leaves the local process.

from appworld_sdk import AppWorldClient

class MyAgent:
    """Implement predict() — the SDK checks this via BaseAgent Protocol."""
    def predict(self, *, screenshot, goal, **kwargs) -> dict:
        # Your model inference here
        return {"action_type": 0, "x": 0.5, "y": 0.5, "text": "", "direction": 0}

client = AppWorldClient(api_key="YOUR_API_KEY", base_url="https://api.appworld.dev")
env_info = client.create_environment(task_ids=["FennoteCreateNote", "FennoteDeleteNote"])

result = client.run_local_eval(
    env_id=env_info["id"],
    tasks=["FennoteCreateNote", "FennoteDeleteNote"],
    agent=MyAgent(),
    max_steps=15,
)

print(f"Success rate: {result.success_rate:.0%}")
print(f"Avg steps: {result.avg_steps:.1f}")

Server-Side Eval

Let the platform handle LLM calls:

from appworld_sdk import AppWorldClient

client = AppWorldClient(api_key="YOUR_API_KEY", base_url="https://api.appworld.dev")
env_info = client.create_environment(task_ids=["FennoteCreateNote"])

result = client.run_eval(
    env_id=env_info["id"],
    tasks=["FennoteCreateNote"],
    model_provider="anthropic",
    model_api_key="YOUR_MODEL_KEY",
    model_name="claude-sonnet-4-20250514",
)
print(f"Success rate: {result.success_rate}")

API Reference

AppWorldClient(api_key, base_url, timeout)

Method	Description
list_apps(**filters)	List available apps
list_tasks(**filters)	List available tasks
get_task(task_id)	Get task details
create_environment(task_ids, name)	Allocate device resources
delete_environment(env_id)	Release resources
make_env(env_id, env_type)	Create Gymnasium env
run_eval(...)	Server-side evaluation
run_local_eval(...)	Client-side evaluation

Exceptions

Exception	HTTP	Meaning
AppWorldError	—	Base exception
AppWorldAPIError	any	Server returned an error
EnvironmentNotReady	503	Device not available
SessionExpired	410	Session closed or timed out

License

Apache 2.0 — see LICENSE for details.