Cua Spaces Puts AI Agents on a Shared Mac Desktop

On October 2, 2026, the San Francisco startup Cua launched Cua Spaces, a free Mac application that gives AI agents a desktop they can share with a person. The announcement arrived as a post on X from the company's @trycua account, and coverage of the launch was published the same afternoon by RuntimeWire's Ryan Merket.

A Desktop Built for Supervision

Cua Spaces centers on a simple idea: the desktop an agent works on should be one a human can actually see. A user can stream the agent's screen, watch it move its own cursor through applications, take over with their own cursor when they want to intervene, and hand control back when the agent is on the right track. The model of interaction treats the person as a supervisor in the room, not an audience member behind glass.

The app is built on Cua Driver, the software layer the company describes as the means by which agents operate desktop applications. That layer matters for software that exposes no programming interface, where the only way in is to click, type, and read the screen the way a person would. By packaging desktop access as infrastructure, Cua is betting that the hardest part of computer-use agents is not the reasoning model but the machine it acts on.

Documentation from the company describes a Space as flexible in where it runs. It can be a local container or virtual machine on the user's own Mac, a cloud sandbox, or a machine the user already owns that has been registered with the system. The first release targets macOS, and the announcement notes that Linux images can run as well.

App Teleportation and the Permission Step

The launch's most talked-about feature is app teleportation. Instead of asking an agent to sign into an application from scratch, a user can move a signed-in session, with its files and state, into an agent's Space. The feature removes the friction of repeated sign-ins, though the company is explicit that a transferred session amounts to a live authenticated foothold for whoever controls the destination Space.

According to the company's teleport guide, a transfer happens only after the user reviews exactly which app, files, and state are involved. Sensitive items require an explicit opt-in, and on a Mac the flow may trigger a Touch ID or password prompt. The company advises teleporting only into Spaces the user owns, and its keyvault guide says the vault keeps sessions encrypted on the Mac while mediating requests for access. KuCoin's flash news summary of the announcement noted the same consent-first framing.

The design points to a wider argument across the agent ecosystem. The question is no longer whether agents can act, but under what supervision and with what boundaries. Cua's answer is to make the permission moment visible and revocable, rather than burying it in a setup screen — a stance that echoes the accountability conversation building around autonomous agents, as covered recently on genznewz.com.

The Team Behind Spaces

Cua was founded by Francesco Bonacci, a former Microsoft researcher, and its research lead is Dillon DuPont. Both previously worked on Windows Agent Arena, a benchmark for agents operating a real Windows environment. In their 2024 paper, the team described more than 150 tasks and reported that their Navi agent completed 19.5 percent of them successfully, compared with 74.5 percent for an unassisted human. The gap between machine and person on that benchmark is part of what makes the supervision story matter.

Bonacci left Microsoft in early 2025 and built Lume, a command-line tool for running native macOS virtual machines on Apple Silicon. He founded Cua in San Francisco that March, and the startup later joined Y Combinator's X25 batch. That history suggests a deliberate focus: less on which model reasons about a task, more on the computer the agent acts on.

Pricing, Licensing, and the Road Ahead

Spaces is free to download and source-available under the FSL-1.1-MIT license. The distinction is worth noting for teams evaluating the product: licensing notes in the repository say Spaces releases become MIT-licensed two years after shipping, while Cua Driver itself remains MIT-licensed today. The setup gives teams room to inspect and eventually modify the code they deploy agents into.

The company has signaled where the business model goes next. Pro and Teams plans are in the works, with shared team machines, session handoff, keyvault sync, and admin controls. The launch version centers on an individual supervising a desktop; the future versions aim at coordinating agents and access across teams, with the governance controls that enterprise deployment demands.

Y Combinator's listing for Cua says the startup has closed a seed round, without stating its size. The launch announcement likewise gives no customer or usage figures. For now, the product is a bet that people will hand agents authenticated work. Whether that bet pays off depends on the desktop being genuinely useful — and on the permission boundaries holding up when a Space is shared or delegated. More coverage of the agent ecosystem can be found on the genznewz.com homepage.