Is computer-use agent fleet infrastructure worth building on in 2026?

Q: What is trycua/cua's 'Computer-Use 2.0'?

It's a five-part toolkit for computer-use agents: Cua Driver (open-source, cross-OS control of native apps on macOS, Windows and Linux), Cua Fleets (a paid hosted product for provisioning isolated cloud desktops), CUA-S1 (small, open-weight specialist models for bounded UI decisions), Lume (local macOS and Linux VMs on Apple Silicon), and Cua Bench (a benchmark and trajectory-export tool for training and evaluating agents).

Q: Is trycua/cua actually free to use?

The Driver, Lume, CUA-S1 and Cua Bench are MIT-licensed and free to run yourself. Cua Fleets, the hosted cloud-desktop service at run.cua.ai, is the paid layer: your code claims a desktop from a maintained pool, and the docs note pools can retain paid capacity after a claim ends, so cost is a real operating variable, not a one-time fee.

Q: Is it worth building your own computer-use agent infrastructure instead?

Only if your requirements are narrow enough that a small dependency isn't worth the risk, or specific enough that the open driver doesn't fit. One Launch HN commenter from NonBioS.ai, an AI software-development shop, said they'd built the same kind of Linux VM orchestration from scratch and called it 'a heavy lift.' For most teams that don't need something bespoke, adopting the open layer costs less than rebuilding it.

trycua/cua just crossed 25,000 GitHub stars adding roughly a thousand in a single day. The open driver underneath is worth adopting today. The paid managed fleet on top of it is a build-vs-buy call, not a default, because of who's actually running it.

What "Computer-Use 2.0" actually ships

The GitHub trending velocity is real; maybe worth building pulled the repo directly and confirmed 25,091 stars, 1,726 forks, and 1,027 open issues as of September 20, 2026, on a project created January 31, 2025. But the star count is the least useful thing to build a verdict on, so here's what's actually in the box: Cua Fleets provisions isolated cloud desktops from a maintained pool, the paid product at run.cua.ai. Cua Driver is the open-source piece, giving an agent CLI, MCP, or SDK access to drive native apps on macOS, Windows, and Linux, with documented connectors for Claude Code, Codex, Cursor, and OpenClaw. CUA-S1 is a family of small, open-weight "System 1" models for bounded interface decisions like which value belongs in a form field. Lume creates local macOS and Linux VMs on Apple Silicon. Cua Bench builds computer-use tasks and exports agent trajectories for training and evaluation.

What's genuinely new versus a single-OS VM tool

The project started in 2025 as an Apple Silicon-only virtualization layer, Lume, built on Apple's Virtualization.Framework. Cross-OS driver support for Windows and Linux is the actual content behind the "2.0" label, not a version-number bump. That matters because most computer-use tooling before this point either lived inside a browser sandbox or was locked to macOS. An agent that needs to drive a legacy Windows desktop app or a Linux CAD tool previously meant building that bridge yourself.

Where the free layer stops and the paid layer starts

Cua Driver, Lume, CUA-S1, and Cua Bench are all MIT-licensed. Run them yourself, for free, today. Cua Fleets is the commercial product: your code claims a desktop from a maintained pool and runs commands against it through the same Sandbox SDK used locally. The documentation includes a detail worth reading before you scale usage: pools can retain paid capacity after a claim ends, so cost is an ongoing operating variable, not a one-time line item. If your plan is to depend on the managed fleet in production, budget it like infrastructure spend, not like a free open-source dependency.

Is this mature enough to build on, or still DIY territory?

Cua AI, Inc. was founded by Francesco Bonacci, went through Y Combinator's Summer 2025 batch, and raised a $500K seed in June 2025. That's a real company with real traction, not a solo weekend project, but it's a small team carrying a large surface area: 1,027 open issues against 25,000-plus stars is a genuine adoption signal and a genuine maturity caveat at the same time. The honest read is that the open driver layer is worth adopting now if it saves you from building your own OS-level automation bridge, and the paid managed layer is worth piloting, not yet worth treating as a guaranteed dependency without your own fallback.

The receipts

On Cua's own Launch HN thread, a commenter from NonBioS.ai, an AI software-development shop, described paying the exact cost this project now removes: "We, at NonBioS.ai [AI Software Dev], built something like this from scratch for Linux VM's, and it was a heavy lift. Could have used you guys if had known about it." That's the pain this layer solves, verified against a real team that already paid to solve it themselves. On the traction side, Cua's own repository, not a secondhand claim, shows 1,898 merged pull requests, which is the more honest maturity signal than a star count: an actively maintained, actively contributed-to project, run by a very small team.

The verdict isn't "wait." It's "adopt the driver, pilot the fleet." A three-person, seed-stage company built something genuinely useful faster than most teams could build it themselves. That doesn't make it enterprise infrastructure yet. It makes it worth a real evaluation before your next quarter's roadmap assumes you'll build the OS-level bridge in-house.

Related: Is a computer-use AI agent worth building in 2026?, on the narrow-vertical-agent wedge this infrastructure supports. Also see do you still need to host your own agent harness in 2026? and is an AI model-routing tool worth building in 2026?

Frequently asked questions

What is trycua/cua's "Computer-Use 2.0"?

A five-part toolkit: Cua Driver (open-source, cross-OS control of native apps on macOS, Windows and Linux), Cua Fleets (a paid hosted product for isolated cloud desktops), CUA-S1 (small, open-weight specialist models for bounded UI decisions), Lume (local macOS and Linux VMs on Apple Silicon), and Cua Bench (a benchmark and trajectory-export tool for training and evaluating agents).

Is trycua/cua actually free to use?

The Driver, Lume, CUA-S1, and Cua Bench are MIT-licensed and free. Cua Fleets, the hosted cloud-desktop service, is the paid layer, and the docs note pools can retain paid capacity after a claim ends, so cost is an ongoing variable, not a one-time fee.

How big is the company behind trycua/cua?

Small. Cua AI, Inc. was founded by Francesco Bonacci in 2025, went through Y Combinator's Summer 2025 batch, and raised a $500K seed in June 2025. The repo has grown to 25,000-plus GitHub stars and 1,700-plus forks on that team size.

What does the open issue count tell you about maturity?

trycua/cua carries 1,027 open GitHub issues against roughly 25,000 stars. That ratio says the project is genuinely used and genuinely still rough in places, so budget for edge cases rather than assuming enterprise-grade polish out of the box.

Is it worth building your own computer-use agent infrastructure instead?

Only if your requirements are narrow or specific enough that the open driver doesn't fit. A Launch HN commenter from NonBioS.ai said they'd built the same kind of Linux VM orchestration from scratch and called it "a heavy lift." For most teams, adopting the open layer costs less than rebuilding it.

What operating systems does Cua Driver actually support?

macOS, Windows and Linux, connected through a CLI, MCP, or typed SDKs, with documented integrations for Claude Code, Codex, Cursor, and OpenClaw. Cross-OS support is the real advance behind the "2.0" label; the project's earlier version was Apple Silicon-only.