Cua
6 Tools · 52 Intel
Cua is a Y Combinator Spring 2025 startup founded by Francesco Bonacci, previously an engineer at Xbox and Microsoft AI, that builds open-source sandboxed VM infrastructure for AI agents that control full desktop operating systems across macOS, Linux, and Windows. The project has amassed over 19,000 GitHub stars and has been adopted by research teams at Google DeepMind, Nous Research, and Snorkel AI.
Is this you? Sign in with X to claim this profile.
Tools
Bootstraps macOS sandboxes for computer-use AI agents that drive real apps in the trycua ecosystem.
HashiCorp Packer plugin that builds macOS VM images with Lume for agent workloads.
Agent-ready desktops on Mac, Linux, or your own cloud, with session teleport and a local Keyvault.
Gives AI agents their own full desktops on your Mac, owned machines, or cloud, with app session teleporting
Sample project showing how to run an agent in a cloud sandbox with the Cua Python SDK.
Intel
Cua Cloud Fleet added declarative configuration for sandbox pools through its SDK and a public trycua/fleets Terraform provider, covering static warm pools and claim-driven autoscaling for reusable Linux and Windows environments.
Cua Driver's new Computer History feature keeps an encrypted local record of which apps and capabilities an agent used and whether actions succeeded, letting a new session reuse a working route instead of re-exploring; it excludes screenshots, keystrokes, and typed text.
Cua Driver 0.28.0+ can now serve its desktop-automation skill (guidance on tool choice, element targeting, focus handling, and action verification across macOS, Windows, and Linux) to compatible agents over a single MCP stdio connection.
Cua open-sourced CUA-S1-FORMS, a specialist model that fills out forms in a single pass by selecting from a fixed action set (FILL, CHECK, CLICK, SKIP), with the full MIT-licensed pipeline for synthetic data, training, and evaluation released alongside it.
Cua built jev-use, pairing TypeSafe's Jev decision model with Cua Driver so a fast classifier selects among bounded action candidates instead of calling a full LLM at every step; in one measured 2048 run this was 5x faster and about 1,000x cheaper than using Astra directly.
Cua's tutorial walks through its cua-agent framework, which replaces the manual OpenAI computer-use control loop from Part 1 with a higher-level abstraction supporting multiple model providers, local models, and a Gradio UI.
Cua adds six cursor motion styles to its open-source Cua Driver for computer-use agents, including comet swoop, spring settle and adaptive, chosen from 82 prototypes. A new signature arc is the default, and the previous motion remains available.
A step-by-step guide to building a Computer-Use Operator on macOS using OpenAI's computer-use-preview model and Cua's cua-computer package, covering VM setup, action handling, and safety checks.
Cua announces a Hermes plugin that creates a clean Space per task on your own Mac, Windows, Linux or cloud VM, runs the work there, and deletes it. Includes install steps and a prompt for checking a PR on Linux and macOS with screenshots from each.
Adds a human/human agent mode and a hybrid AI-plans-human-executes pattern to Cua's Agent framework, with a web UI for taking over an agent mid-task, useful for generating training demonstrations or handling low-confidence fallbacks.
Cua launches Cloud Sandbox, letting developers spin up pre-configured, browser-accessible desktop environments for computer-use agents in seconds instead of managing local macOS VMs, with pay-per-use pricing and BYOK model support.
The Cua CLI adds terminal commands to create, list, and connect to Linux, Windows, or macOS cloud sandboxes, with browser-based OAuth login, .env export for SDK integration, and both flat and grouped command styles.
Cua built a process-scoped Metal capability shim that unlocks newer GPU code paths inside a Lume macOS VM, making llama.cpp run 7-16x faster on models like TinyLlama, Gemma 4, and Meta's Muse Glimmer 30B on Apple Silicon, releasing source and benchmarks openly.
Cua shipped a stable driver giving agents a synthetic cursor built into Omarchy's Hyprland compositor, letting an agent work a background window while the user's own pointer stays free, built with Omarchy's DHH and collaborators.
Cua demoed a technical preview of background computer-use on Linux/Wayland, testing two compositor-side isolation paths (direct wl_pointer delivery vs. a separate wl_seat) that keep agent input from disturbing the user's foreground session.
Breaks down how cua-driver 0.5.7 controls Linux desktop apps in the background for agents like Claude Code or Codex: AT-SPI2 over D-Bus for accessibility trees, XTEST for synthetic input, and a separate painted cursor, with per-toolkit quirks across GTK, Qt, and Electron.
Cua Cloud Fleets now offers a managed Omarchy (Arch Linux + Hyprland) cloud desktop, booted as a full KubeVirt VM rather than a themed container, controllable via the Cua Sandbox SDK or Cua Driver MCP.
Cua Driver 0.19.0 adds an extension-free bridge between a native OS window and its exact Chromium tab via CDP, letting an agent use semantic in-page browser actions and native desktop control in one session without installing a Chrome extension.
OpenClaw 2.0 integrates Cua Driver natively rather than through an external MCP connection, owning the agent tool, session lifecycle, and action contract, and can be launched directly as a desktop fleet on Cua Cloud Fleets.
Cua Cloud Fleet added signed URLs that grant temporary, revocable access to a single sandbox service (1 minute to 24 hours), created and revoked through the TypeScript Fleet SDK without keeping the underlying sandbox claim alive.
Cua Fleets now lets you publish custom Linux or Windows VM images to a registry and boot agent sandboxes from them by digest, similar to Docker image workflows.
Introduces CuaBot, a CLI giving agents like Claude Code, Codex, or Aider a sandboxed Linux desktop with computer-use tools, streamed via Xpra so the agent's cursor and windows coexist with the user's on the same screen instead of taking over control.
Cua Cloud Windows Sandboxes reach general availability with sub-second hot starts and three sizes starting at 8 credits/hour, alongside an invite-only preview of Cloud macOS Sandboxes running on bare-metal M1/M2/M4 hosts.