Show HN: CUA-S1 – A System One Model for Computer Use

How to use it

Each component installs and produces a first result independently:

• Cua Driver (macOS/Linux): /bin/bash -c "$(curl -fsSL https://cua.ai/driver/install.sh)"; Windows: irm https://cua.ai/driver/install.ps1 | iex. First result: connect an agent, have it compute 6×7 in Calculator and verify the display reads 42. Connect via CLI, MCP, or typed SDKs. Background delivery can drive apps without moving your pointer or stealing focus when the app and platform support it.
• Lume (local macOS/Linux VMs on Apple Silicon via Apple's Virtualization.Framework): /bin/bash -c "$(curl -fsSL https://cua.ai/lume/install.sh)".
• Cua-Bench: needs Python 3.12/3.13 and uv: `` uv tool install 'cua-bench[browser]' uv tool run --from 'cua-bench[browser]' playwright install chromium `` First task runs simulated — no VM, Docker, or model API key required.
• Cua Fleets: provision cloud desktops at run.cua.ai. Note: pools can retain paid capacity after a claim ends, so the cleanup steps matter for cost.

Local sandboxes and Fleets share the Sandbox SDK, but credentials, images, operations, and runtime requirements differ.

Agent integrations exist for Claude Code, Codex, Cursor, and OpenClaw.

Limits

• CUA-S1 is a source-only research release on GitHub (Python model code, synthetic-data generation, training, eval); model weights live separately on Hugging Face (CUA-S1-FORMS). Only the first "forms" profile exists. The authors explicitly frame "System 1" as an engineering analogy for fast bounded decisions — not an architecture claim or a replacement for a general agent's planning; application code still orders the actions.
• Licensing is not uniformly MIT. The core is MIT, but cua-perception is installed separately and is not MIT: its reviewed candidate bundles an AGPL-3.0-only OmniParser artifact, Apache-2.0 PP-OCR artifacts, and a separately packaged ONNX Runtime. The page warns that **redistributing it or offering it to users over a network can trigger AGPL-3.0 source obligations** — Cua cannot relicense the detector. cua-som is likewise AGPL-3.0-or-later (with an Ultralytics dependency retaining its own license). Check the perception third-party notices before installing, hosting, or redistributing.
• Verify each model/dataset card and the exact OmniParser model revision for artifact-specific terms before redistribution.

For reproducibility the project asks you to cite the specific release or commit (CITATION.cff provided).