#1010th of 18 in Agent backends
openai-cua-sample-app
Computer-use agent loops for Playwright browsers and PyAutoGUI desktops
- Stars
- 1.9k
- License
- MIT
- Last commit
- Sep 2026
- Language
- TypeScript
Overview
Two agent loops on the OpenAI Responses API where the model writes code against a persistent runtime: a TypeScript agent driving a browser through Playwright and a Python agent driving the real desktop through PyAutoGUI. A shared console on port 3000 runs scenarios against bundled lab apps and records screenshots and replay JSON. For developers building computer-use agents who want the loop, not a product.
Who it is for: Developers building computer-use agents from the loop up
Strengths
- Lab apps and replay traces let you test the loop without touching real sites
- Both agents share one console and contract types; tests in each app
- Persistent execution worker pattern reduces model round trips
Weaknesses
- No sandbox: generated code runs with your user permissions
- Python agent controls your real mouse and keyboard
- OpenAI-only; requires model access for computer use
- Pinned Node 22.20.0 and pnpm 10.26.0
What it needs
- TypeScript, openai
- Needs openai-api-key, playwright-chromium, uv
- env example file
Also in Agent backends
See all 18| Rank | Project | Score |
|---|---|---|
| 1 | eve-software-factory-templateeve pipeline that turns GitHub or Linear issues into reviewed draft PRs | 65 out of 100 |
| 2 | marketing-team-eve-templateeve lead agent delegating to five marketing specialists with approval gates | 59 out of 100 |
| 3 | personal-agent-templateeve and Nuxt personal agent with Slack, GitHub, Linear and per-user memory | 57 out of 100 |
| #4 | adk-recipesRunnable Agent Development Kit recipes, from single patterns to deployable agents | 53 out of 100 |
| #5 | ai-townGenerative-agents town simulation on Convex with Ollama by default | 52 out of 100 |