xlang-aiOfficial·17 Agent Skills Included

CUA-Gym

Generate verifiable training tasks for computer-use agents

Generates verified training tasks, environment states, and reward functions for computer-use agents across 110 desktop and web environments. Eliminates hours of manual task authoring by running generator, discriminator, and orchestrator agents that produce tested reward triples automatically. Ships a 32,000-task dataset plus per-app guides covering LibreOffice, Chrome, GIMP, Blender, Grafana, and mock web apps.

185|14

↗

npx skills add xlang-ai/CUA-Gym --all -g -y

All Skills in This Repository (17)

Pure Emerald Level Indicators

📦 In Repo

xlang-ai

penpot

Automate Penpot project setup and design state via HTTP APIs.

libreoffice-writer

Create, modify, and verify LibreOffice Writer .docx files with Python.

openshot

Automate OpenShot video project setup and verification with Python.

libreoffice-calc

Create, modify, and verify .xlsx files with Python's openpyxl library.

overleaf_reward

Generate reward.py scripts that verify Overleaf LaTeX project states via HTTP APIs.

pdf

Generate, modify, and verify PDF files using Python libraries.

grafana

Automates Grafana dashboard setup, manipulation, alerting via HTTP APIs for CUA-Gym tasks.

excalidraw

Automate Excalidraw scene setup and verification for RLVR environments.

vlc

Automate VLC media player configuration, playback control, and verification via Python.

mock_websites

Set up and manipulate mock web application state via HTTP APIs.

drawio

Manage draw.io diagrams programmatically via Playwright CDP and Python XML.

chrome

Automate Chrome browser state setup and verification with Playwright.

Official

Advanced

Frequently Asked Questions

FAQPage Schema

How to install CUA-Gym?▼

Run `npx skills add xlang-ai/CUA-Gym --all -g -y` in your terminal to install all skills in this suite globally.

What does CUA-Gym do?▼

It automatically creates verified training tasks for computer-use agents, bundling each task instruction with an executable environment state and a programmatic reward function.

Which applications does CUA-Gym cover?▼

It covers 110 environments including LibreOffice, Chrome, VS Code, GIMP, Blender, VLC, Grafana, Penpot, and 94 mock web apps like Gmail, Slack, and Notion.

How are the generated tasks verified?▼

A discriminator agent writes reward scripts without seeing the generator's code, and an orchestrator iterates until the reward scores 1.0 on the golden state and 0.0 on the initial state, followed by a majority-vote filter.

Can I download the pre-built CUA-Gym dataset?▼

Yes. The full dataset of 32,112 verified task bundles is available on Hugging Face at xlangai/CUA-Gym and can be loaded with the standard datasets library.