1/ We're reimagining what it means for your agents to work with all your computers. Today we're excited to share Cua Spaces, built on Cua Driver and our virtualization stack, rolling out on macOS today. Download it now, it's free and source-available: spaces.cua.ai

cua-driver is the way. Making Claude use Codex stuff is a recipe for flakiness and things breaking on updates IMO. cua has been insanely fast and stable for me. Hyped for cua-spaces too! Hope it comes to windows & linux too since I run all 3 OSes.

my cal is open for computer use collison-installs! cua.cal.com/f/cua-collison-i…

cua-driver serve --permission-mode unrestricted

Glad you all like Spaces. You'll love what's next 👀

🤯🤯

one of the most frustrating things is watching people holding up “look at this one shot blender thing this model did oh my god everything is solved!” only to then not have that same model be able successfully log into something or some other simple task. Computer use, while often “slower”, is going to be an important part of various workflows and solutions stacks and the cua team is really cooking

Federico Viticci spent this week running Claude, inside T3 Code, to orchestrate a second Mac over Tailscale, and that's a more interesting story than Cua's own announcement post let on. Cua Spaces shipped free and source available on macOS on October 2, built on Cua Driver, the MIT licensed background driver that lets an agent click, type, scroll and read the accessibility tree on a native desktop app without hijacking your cursor. Driver runs as an MCP stdio server, a long running daemon, or a one shot shell command, and its GitHub community is already past 27,000 people. For teams that want cloud capacity instead of their own hardware, Cua also sells Fleet, elastic pools of machines priced by the vCPU hour, for training and evaluating agents at scale. What's actually new with Spaces is the mesh. Instead of one agent controlling one machine, Cua is treating every computer you own as claimable capacity on your own network. Viticci's setup paired an M3 Ultra Mac Studio with 512GB of RAM running one agent against an M5 Ultra on 256GB running another, with Claude as the coordinator between them. This sits beside the existing computer use stack, not on top of it. Anthropic ships its own computer use API, OpenAI has Operator, Browserbase and E2B sell the sandbox layer most agents run inside. Cua's bet is that the interesting problem isn't any single sandbox, it's treating every machine you already own as fleet capacity. Integrations exist today for OpenAI, Qwen Code and Factory Droid. Orchestrating across hardware is the easy half of this. Nobody's shipped the part where two agents on two machines disagree about what the task actually is. x.lingyaoai.com/trycua/status/21060571…

Wow this is incredible. Agents can each get their own “space” to work so it’s separated from you (and other agents), each getting their own cursor etc.

I use CUA Driver daily here is how: >100% REPLACING CODEX CUA >2X cheaper Parallel Competitor analysis >Scraping with human like behaviour Do this and your coala bro simply can execute with no maintance I won't let frontier models to parse all my data. (No DOTs were used) So I've cooked this animation in matter of CUA slight rebranding You met me at very Coala time in my life @francedot How do you like it?

🔥

This. is. Wild. Source available.... recommend (continuous) review prior to use for something this powerful.

Okay so this is pretty wild, here's what I've done: Set up a Cua Space on the M3 Ultra Mac Studio with 512 GB of RAM, with @t3dotcodes running on the M5 Ultra with 256 GB. Claude in T3 Code can now orchestrate computer use on another Mac on my Tailnet. There are currently 8 Xcode simulators running on the second Mac, so I can free up RAM for local AI on the M5 Ultra. It rocks. The @trycua team keeps impressing me.

出典: Cua の投稿（10/2）と、Cua Spaces の公式ページ

Give your agents their own desktop: Cua Spaces is free on macOS 26+ (Apple silicon), no account needed. You and every agent get your own cursor, and Teleport moves a logged-in app session into it. Source-available under FSL — fine for anything except a competing product.

Computer-use agents keep hitting the same wall, and it isn't a benchmark. It's a permission dialog nobody taught the agent to handle, or a login session that evaporates the moment the agent switches apps. Cua Spaces, out this week from YC backed Cua AI, treats that as the actual product problem. A Space is a full virtual desktop, on your Mac through Lume's Apple Silicon VMs, on machines you own, or in the cloud, pre-configured with the permissions and tools an agent needs so it stops stalling at dialogs. Cua Driver gives the agent computer vision and input control over real apps instead of an API shim. RCDP, their remote desktop protocol running over QUIC, streams the desktop while keeping cursors separate, so you can watch an agent work without fighting it for the mouse, and multiple agents can share a machine without colliding. "Teleport" moves a whole session, tabs, cookies, signed in state, into a Space instead of making the agent log in from scratch. The core driver is MIT licensed, sitting at 23.1k GitHub stars and 1.6k forks. Everything is free during early access. It's entering a sandbox market that already has real money in it, just not for this exact job. E2B raised a $21M Series A and has started over a billion sandboxes for customers like Perplexity and BlackBox AI, pricing around five cents a vCPU hour. Daytona closed a $24M Series A in February on sub-90ms boot times. Modal raised $355M at a $4.65B valuation in May. All three are selling code execution: spin up a container, run a script, tear it down. None of them are selling a desktop an agent can see and click through, which is a different problem with different failure modes. The part that's actually hard here was never the compute. E2B and Daytona solved boot time years ago. What's unsolved is the boring stuff: permission state, session continuity, who owns the cursor when three agents share a screen. That's a UX problem wearing infrastructure clothes. x.lingyaoai.com/trycua/status/21060571…

Things are getting so cool

one thing about spaces that i'm excited about is being able to easily share compute between friends. i'm a big fan of making the most of existing compute, instead of spinning up more and more sandboxes @cognition's RSA-260 factorisation used spare compute across their clusters at "no marginal cost". while the technicals are admittedly nuanced, the idea remains the same in maximising usage

Oh wow, this one is really good!!

Finally, a proper place for agents to live and work. Looks well made.

Agents have their own space, on your machine, VM, server. Nice CUA