Cua Driver opens macOS background app control with multi-cursor support for Claude Code and Codex
Cua Driver open-sourced a macOS driver that lets agents control apps in the background with multi-player and multi-cursor support. It matters because it turns background computer use from an app-specific feature into a reusable primitive that any agent loop can adopt.
TL;DR
- trycua's launch thread says Cua Driver is an open source macOS driver that lets agents, including Claude Code, Codex, and custom loops, control apps in the background with multi-player and multi-cursor support.
- According to trycua's implementation note, the background-control trick uses SkyLight private APIs plus yabai's focus-without-raise pattern, and Cua framed it as an open alternative to Codex's OpenAI-only background computer use.
- trycua's dev-loop demo shows Claude Code reproducing a visual bug, writing a fix, and QAing it while the agent harness stays frontmost.
- trycua's demo-capture clip adds a second workflow, agent-driven product demo recording with zoom-on-click export, while a later retweet from trycua says the team replaced Screen Studio for some internal demos.
- trycua's off-screen Chrome demo and trycua's Messages demo push the same primitive beyond coding, into pulling visual context from tabs you never raise and sending replies without switching Spaces.
You can browse the repo and the linked macOS window internals writeup. The pitch is unusually concrete: one demo keeps the coding harness frontmost while the agent fixes a UI bug elsewhere, another turns the driver into an automated demo recorder, and a third sends a Messages reply without stealing focus.
Cua Driver
Cua is packaging background computer use as a reusable macOS primitive, not as a feature trapped inside one agent product. In the launch post, trycua says any agent can use it, naming Claude Code, Codex, and custom loops.
That matters because most desktop agent demos still assume the agent owns the foreground. Cua's claim is the opposite: the human keeps their current app and the agent works elsewhere.
The official pointers are the GitHub repository and a linked inside macOS window internals post, both surfaced in trycua's final thread post.
Window internals
The most useful technical detail in the launch thread is the implementation stack:
- SkyLight private APIs
- yabai's focus-without-raise pattern
- a background-control UX that trycua credits Sky for proving out first
In that post, trycua explicitly contrasts the result with Codex's background computer use from the prior week, calling Codex's version OpenAI-only and Cua's version open source. That makes this feel less like a shiny wrapper and more like a missing systems layer for macOS agents.
Background workflows
The dev-loop clip in trycua's demo is the cleanest coding example: Claude Code reproduces a visual bug, writes the fix, and runs QA while the controlling harness never loses the foreground.
The second workflow in trycua's recording demo uses the same driver for delegated demo creation. trycua says the agent drives the app while cua-driver record exports the session with zoom-on-click.
That workflow was not just marketing copy. In a later retweet, trycua highlighted an internal use case where the team canceled Screen Studio for some demos and used cua-driver instead.
Off-screen context
Background control also changes what counts as available context. In trycua's Chrome example, Claude Code pulls visual information from a YouTube scene inside a tab that was never raised.
The last demo pushes the idea into personal-assistant territory. trycua's Messages clip shows an agent sending a Messages reply while the user keeps typing in their editor, with no Space switch during the handoff.