Cua – Scale Computer Fleets for AI Agents

TL;DR

TL;DR: Cua is an open-source platform (MIT licensed) that provisions Linux, Windows, macOS, and Android VMs as agent sandboxes, with an MCP server and CLI for driving computer-use agents at scale.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: cua.ai
• Source repository: github.com/trycua/cua — README read 2026-07-17
• License: MIT (verified via GitHub API license.spdx_id)
• HN launch thread: news.ycombinator.com/item?id=43773563 (172 points, YC X25)
• Related Lume Show HN: news.ycombinator.com/item?id=46670181

What Is Cua?

Cua (YC X25) is open-source infrastructure for building, benchmarking, and deploying AI agents that control full desktops. It solves the “computer use” problem end-to-end: provisioning real VMs across four operating systems, exposing them via MCP and CLI, and providing evaluation harnesses to measure whether agents completed tasks correctly.

The README describes it as:

Open-source infrastructure for Computer-Use Agents. Sandboxes, SDKs, and benchmarks to train and evaluate AI agents that can control full desktops (macOS, Linux, Windows).

The platform has four main components:

• Cua Driver — background computer-use for macOS, Windows, and Linux. Agents click, type, and verify without stealing cursor or focus. Ships as an MCP stdio server.
• Cua Sandbox — on-demand VM provisioning for Linux, Windows, macOS, and Android. Warm pools for parallel batch evaluation.
• Cua Bench — evaluation and gym authoring for cross-surface agent benchmarks.
• Lume — macOS VM management on Apple Silicon, built on Virtualization.framework.

Setup Workflow

Prerequisites

• macOS, Linux, or Windows host
• 8GB+ RAM recommended
• Docker (for cloud sandbox), or Virtualization.framework (macOS, for Lume)

Install the Cua Driver

macOS / Linux:

/bin/bash -c "$(curl -fsSL https://cua.ai/driver/install.sh)"

Windows (PowerShell):

irm https://cua.ai/driver/install.ps1 | iex

Wire into Claude Code

Once installed, configure Claude Code to use Cua as an MCP server by adding this to your Claude Code config:

{
  "mcpServers": {
    "cua": {
      "command": "cua",
      "args": ["driver", "run"]
    }
  }
}

Run a Cloud Sandbox

For larger workloads, provision a cloud sandbox:

cua sandbox create --os linux --region us-east-1
cua sandbox connect <sandbox-id>

Use the CLI Directly

# List available sandboxes
cua sandbox list

# Run a one-shot command
cua run --os macos --image ghcr.io/trycua/macos-base "python3 /tmp/agent.py"

Deeper Analysis

Cross-platform consistency is Cua’s core value proposition. Rather than maintaining separate automation stacks per OS, you write once against the MCP interface and run on Linux, Windows, macOS, or Android. The Driver handles OS-specific input routing (macOS Accessibility API, Windows UI Automation, Linux X11/Wayland).

For model training, Cua Bench provides eval harnesses where you define a task, point the system at a warm pool of VMs, and collect trajectory data. Human-reviewed golden trajectories are available for selected benchmark scopes.

Pricing is free for open-source use locally. The GitHub repo is MIT licensed. Dedicated hosted fleets and BYOC infrastructure are available by contacting the team — no self-hosted server component required for evaluation.

Practical Evaluation Checklist

• [ ] Driver installs without errors on target OS
• [ ] MCP server starts and responds to stdio ping
• [ ] Agent can click/type in a headless macOS VM via Driver
• [ ] cua sandbox create provisions a VM within 60 seconds
• [ ] Trajectory data exports in structured format (JSON)

Security Notes

• Cua Driver on macOS requires Accessibility permissions (for UI input)
• Windows Driver uses UI Automation — no admin required for standard apps
• SOC 2 Type I certification in progress for hosted infrastructure
• BYOC (Bring Your Own Cloud) option keeps data on your own infrastructure

FAQ

Q: How is this different from Docker-based containers? A: Docker containers can’t simulate full desktop interaction (clicking buttons, reading screen pixels, using native macOS/Windows apps). Cua runs real VMs with full OS kernels, so agents interact exactly as a human would.

Q: Does it work with Cursor, Codex, or OpenClaw? A: Yes. Any tool that supports MCP stdio servers can connect to Cua Driver. The same CLI works across all three.

Q: Is there a hosted option? A: Yes — cloud sandboxes are available by request. You can also run everything locally with the open-source build.

Q: What Apple Silicon macOS VMs does Lume support? A: Lume 0.2 builds on Apple’s Virtualization.framework (macOS 13+). It supports macOS VMs with unattended setup for eval and training workloads.

Conclusion

Cua tackles the full computer-use agent stack in one open-source project: VM provisioning, background input routing, MCP interface, and evaluation benchmarks. If you’re building or evaluating agents that need to drive real desktops — not just call APIs — Cua is worth a look. MIT licensed, with hosted cloud options for teams that need scale without managing infrastructure.