OpenAI Bilgisayar Kullanan Temsilci Örnek Uygulaması

Computer-use agents are AI systems that operate software to complete a task. They inspect an interface, choose an action, execute it, and check the result. At OpenAI, we build this loop around models that write code to interact with software. Code lets the model combine actions, process observations, and choose when to look again.

A persistent runtime keeps useful state and helper functions available between calls. The model can write a loop to fill several fields, check that each change took effect, and return only the text or screenshots it needs. This can reduce model round trips and repeated input context while giving the model feedback to correct mistakes.

You need Node.js 22.20.0 and an OpenAI API key with access to the configured model. Corepack uses the repository's pinned pnpm 10.26.0 . Run these commands in your terminal, replacing the API key placeholder:

Both apps read the root .env ; shell variables take precedence. You can save the key in .env to use it in later sessions.

Run either quickstart below from the repository root. Once it starts, open the console , choose a scenario, edit its prompt, and select Start Run . Runs make real API calls. Use Stop to interrupt a run; use Ctrl+C and wait for shutdown before switching apps.

The Python agent runs its server, agent loop, and persistent PyAutoGUI worker in Python, controlling a visible browser on your desktop.

Install uv , use Python 3.10+ (the repo pins 3.12), and follow the Python README for desktop permissions, platform setup, and a code walkthrough.

Each run gets a fresh copy of a lab template. Workspaces, screenshots, and replays stay in the selected app's ignored data/ directory. Use Replay JSON to inspect the recorded trace and the screenshot timeline to review earlier environment states. Generated dependencies, builds, and caches are omitted above. See the lab guide for task examples.

Run finished means the agent loop and cleanup ended normally. Inspect the screenshots, recorded trace, and model response to judge whether the requested task was accomplished.