OpenGCUA

OpenGCUA: Open source research stack for generalist computer use agents combining models, environments, and cross-device SDKs.

OpenGCUA is an open project that develops the full stack for generalist computer use agents. The system allows models to complete diverse computer tasks through visual interaction and programmatic access. The model decides how to observe the environment using screenshots, accessibility trees, command outputs, or API responses. It acts through GUI operations, shell commands, API calls, or programs, and organizes its work across steps. Tested with frontier models including GPT-6 Astra, Claude Opus 5, Gemini 3.8 Flash, GLM-5.3 Flash, Qwen3.8 Max, Qwen3.8 Flash, Kimi K3, and DeepSeek-V4.1 Flash, the project provides evaluation trajectories and benchmarks.

OpenGCUA at a glance

Best for: Researcher
Runs on: open source
Pricing: Free
Released: October 2026
Category: AI agent tool
Website: tongyi-mai.github.io

Key features

Choose observation types dynamically
Execute GUI operations and shell commands
Run code with control flow
Track progress and unverified changes
Access cross-device SDKs and benchmarks
Utilize rolling screenshot buffers
Release execution trajectories

Use it for

Run computer use agents on local machines
Evaluate frontier models on OSWorld-v2
Combine GUI clicks with shell commands
Automate tasks across multiple interfaces

Alternatives to OpenGCUA

OpenResearch — Run parallel research workflows locally or on remote compute.
Cua — Scales computer fleets across four OS families with one MCP/CLI driver.
BrowserOS — Runs entirely locally and lets AI agents use real user logins to complete tasks.
Agent Reach — Provides zero API fees, multi-backend routing, and local credential storage.
GPT Researcher — Ranked #1 on Carnegie Mellon's DeepResearchGym benchmark for academic research.
GOD — Real-time control and intervention for agent simulations.
CAMEL — Builds complex multi-agent societies for data generation and simulation.
MS-Agent — It supports multimodal inputs, autonomous scheduling, and modular tool integration.
ART — Enhances LLM agent performance and reliability using reinforcement learning techniques.
OWL — Ranks first among open-source frameworks on the GAIA benchmark with a score of 69.09.
GEAR — Organizes recursive self-improvement through evolutionary search and reproducible evaluation.
Fara — It operates visually to perform multi-step web tasks without accessibility trees.

All OpenGCUA alternatives