Pinned Tweet
1/ Today we're launching Cua Driver Bench, together with Scale AI's AgentEnv.
Can an agent do a desktop task in the background while you keep working undisturbed? On macOS, Cua Driver fully passed 58%, Codex computer use 33%, Vercel's agent-desktop 20%, browser-use 3%.
Cua retweeted
Use @trycua and catch me if you can
United Airlines is blocking personal AI assistants.
Plenty of privacy, fraud & safety concerns, surely.
But I can’t help but feel they’re trying to hinder an automated way of finding (or rebooking) the cheapest fare before widespread use. united.com/en/us/newsroom/an…
5
301
Cua retweeted
Today I'm releasing Cupola. The coding agents on your Mac show up on your iPhone as chats in one inbox: Claude Code, Codex, T3 Code, OpenCode and Cursor.
Reply to them, approve what they ask for, and see your Mac's screen live from the phone via Wi-Fi and Tailscale.
Use @trycua to control your computer as well and see what's happening directly on you phone!
Cua retweeted
周囲のプログラマ、特定AI企業にロックインされたくなさすぎて軒並み Cua Spaces + Composio 使ってオレオレ Dots/Muse/Duet を自作し始めてる。 パーツは既に揃ってるので多分すぐコモディティ化してくれると思う
Cua retweeted
7 CLI tools you HAVE to give to your AI agents:
1. Notion CLI by @NotionHQ
I've been using Notion for years, so you can be sure I am abusing this.
You can let your agent read, create, update, and organize pages directly from the terminal.
Incredible for people building a secondbrain in notion.
10/10
2. Whop CLI by @whop
Been a fan of Whop way before they went mainstream, been using the CLI as soon as it dropped.
If you don't know what Whop is, it's a platform that lets you sell products.
With the Whop CLI, your agents can manage your businesses on Whop for you.
They can create products, build websites, manage customers, pull analytics, and more.
10/10
3. Hyperframes by @HeyGen
One of the best ways to create and edit videos with agents.
This + claude is TOP TIER.
10/10
4. Agent-browser by @vercel
Browser automation is the future, so make sure you give your agents that.
You can basically create persistent Chrome profiles, and connect to your existing browser.
The team at Vercel is obv cracked, and agent-browser is amazing.
10/10
5. Runware CLI by @runware
Been looking for the cheapest ways to generate images and found this gem.
Runware lets you generate images, videos, and audio using hundreds of AI models.
With the CLI, you can let your agent batch-generate content for you. This is insane for those who are heavy into AI content.
One of the cheapest options I've found is Z-Image Turbo, at around $0.0026 per image, depending on the settings.
$1 ≈ 384 images.
6. TinyFish CLI by @Tiny_Fish
The best free web search and fetch in the market rn.
I'm a huge fan of this.
I have it plugged into all my agents (Hermes, ChatGPT, OpenCode, OMP, etc.).
Search and Fetch are free hopefully forever lol.
10/10
7. Cua-driver by @trycua
Computer-use is literally one of the best things you can use your agents for.
You can control actual desktop apps with your agent like blender, etc.
Prob the best computer-use tool right now.
And they launched Spaces, which is insane. Idk what the team is on but I want some of it lol.
10/10
Thinking about doing a guide for each, what do you guys think?
Cua retweeted
Cua가 Scale AI랑 같이 Cua Driver Bench 출시.
계속 일하는 중에 AI가 백그라운드에서 데스크탑 작업 잘하나 테스트한 거임
맥에서 Cua Driver 58% 통과
Codex computer use 33%
Vercel agent-desktop 20%
browser-use 3%
사람이 계속 타이핑하는데
에이전트가 키보드 포커스 안 뺏고
포인터도 안 움직이고 일 끝내야 통과
아직 수치가 많이 낮군요
Cua retweeted
everybody better than cua-driver until it’s time to be better than cua-driver
I tested cua-driver 0.38.0 against Vercel agent-desktop 1.2.0 on the same Mac.(@vercel_dev @trycua) For each, I reset Calculator and pressed 4 then 2, and typed AB-TRIAL-042 into a new, unsaved TextEdit document. Both got the exact values while Hermes stayed frontmost at the before/after checks. That is one attempt per task, not a speed or reliability benchmark. I captured both sets of app screenshots through CUA because Vercel lacked Screen Recording permission.
For a wider test, Scale AI ran CUA-authored tasks and grading on macOS 26, holding the model, prompts and limits fixed. A full background pass meant finishing through the GUI without disturbing someone typing in front. With Claude Sonnet 5.5, CUA passed 23/40 versus Vercel's 8/40; with Haiku 5.5, it was 18/40 versus 10/40. My two-task check was a tie. CUA wins that published benchmark, not every desktop task.
For CUA, I'd work on a real background drag path and stop its brief pop-up menus from appearing over the user's window. I'd also fix Hermes's element-ref wrapper: it rejected my button click, forcing a coordinate fallback. For Vercel, I'd detect AX actions that shift focus, keep them out of background mode, and make stale-ref refresh cheaper than another agent turn. I'm keeping CUA as my default for background Mac work.
Im gonna link the trial data under here. But all in all I think soon computer use will be fine tuned and better than any of us lol.
Also mind you. This took me like an evening so if there was a problem in the way I conducted the comparison please let me know and I will retry it! I was doing this just for fun and to learn
Cua retweeted
1/ love seeing independent tests like this!
even as we've launched more products under Cua, we haven't slowed down on Cua Driver.
as of 0.44 now: faster typing, several steps batched into one call, and a lot of fixes. go and try it out now!
I tested cua-driver 0.38.0 against Vercel agent-desktop 1.2.0 on the same Mac.(@vercel_dev @trycua) For each, I reset Calculator and pressed 4 then 2, and typed AB-TRIAL-042 into a new, unsaved TextEdit document. Both got the exact values while Hermes stayed frontmost at the before/after checks. That is one attempt per task, not a speed or reliability benchmark. I captured both sets of app screenshots through CUA because Vercel lacked Screen Recording permission.
For a wider test, Scale AI ran CUA-authored tasks and grading on macOS 26, holding the model, prompts and limits fixed. A full background pass meant finishing through the GUI without disturbing someone typing in front. With Claude Sonnet 5.5, CUA passed 23/40 versus Vercel's 8/40; with Haiku 5.5, it was 18/40 versus 10/40. My two-task check was a tie. CUA wins that published benchmark, not every desktop task.
For CUA, I'd work on a real background drag path and stop its brief pop-up menus from appearing over the user's window. I'd also fix Hermes's element-ref wrapper: it rejected my button click, forcing a coordinate fallback. For Vercel, I'd detect AX actions that shift focus, keep them out of background mode, and make stale-ref refresh cheaper than another agent turn. I'm keeping CUA as my default for background Mac work.
Im gonna link the trial data under here. But all in all I think soon computer use will be fine tuned and better than any of us lol.
Also mind you. This took me like an evening so if there was a problem in the way I conducted the comparison please let me know and I will retry it! I was doing this just for fun and to learn
Cua retweeted
I tested cua-driver 0.38.0 against Vercel agent-desktop 1.2.0 on the same Mac.(@vercel_dev @trycua) For each, I reset Calculator and pressed 4 then 2, and typed AB-TRIAL-042 into a new, unsaved TextEdit document. Both got the exact values while Hermes stayed frontmost at the before/after checks. That is one attempt per task, not a speed or reliability benchmark. I captured both sets of app screenshots through CUA because Vercel lacked Screen Recording permission.
For a wider test, Scale AI ran CUA-authored tasks and grading on macOS 26, holding the model, prompts and limits fixed. A full background pass meant finishing through the GUI without disturbing someone typing in front. With Claude Sonnet 5.5, CUA passed 23/40 versus Vercel's 8/40; with Haiku 5.5, it was 18/40 versus 10/40. My two-task check was a tie. CUA wins that published benchmark, not every desktop task.
For CUA, I'd work on a real background drag path and stop its brief pop-up menus from appearing over the user's window. I'd also fix Hermes's element-ref wrapper: it rejected my button click, forcing a coordinate fallback. For Vercel, I'd detect AX actions that shift focus, keep them out of background mode, and make stale-ref refresh cheaper than another agent turn. I'm keeping CUA as my default for background Mac work.
Im gonna link the trial data under here. But all in all I think soon computer use will be fine tuned and better than any of us lol.
Also mind you. This took me like an evening so if there was a problem in the way I conducted the comparison please let me know and I will retry it! I was doing this just for fun and to learn
Cua retweeted
🐚 Introducing SOTA open source ardia CLI
So much for pacing the frontier huh
give claude code, codex, cursor or opencode ur real browser, with ur logins, running in background tabs so ur screen stays untouched
• SOTA Browser use
• per-site notes ur agent remembers, logins kept in keychain
• ardia computer: drives ur native mac apps too, not just the browser powered by @trycua
• hooks n events: run stuff when tabs open, downloads land, or ur idle
• userscripts, themes, config, history n pinned tabs, all from the terminal
on benchmarks it beats aside n opencode browse on tokens with both codex n claude code, n it's fastest on codex
detection-resistant at the engine level, with an optional humanized mode for mouse n typing. it's off by default so it stays fast.
github.com/mantrakp04/ardia
Cua retweeted
can an agent use your Mac in the background without stealing your mouse
@trycua built a bench for it with Scale AI's AgentEnv, the user keeps typing in front while the agent works behind
> cua-driver 58% full passes on macOS
> Codex Computer Use 33%
> Vercel agent-desktop 20%, browser-use 3%
> no driver can hover a background window without moving your real pointer
github.com/trycua/cua/blob/m…
Cua retweeted
HERMES CAN USE YOUR COMPUTER WHILE YOU KEEP USING IT.
Are you letting Hermes handle desktop tasks while you keep working?
You keep writing in Word. Hermes works through export settings in another app or prepares a draft in Outlook.
Same machine. At the same time.
▸ Your cursor stays where you left it
▸ Your keyboard focus stays with you
▸ Your desktops and Spaces don't switch
That's Computer Use, built into Hermes. It drives native apps in the background on macOS, Windows and Linux.
And it works with any tool-capable model, including a local one.
Use it for native UI tasks without a usable API: desktop menus, settings, dialogs and app controls.
Web pages still go through Hermes' browser tool. Files through its file tools. Commands through the terminal.
Turn it on in hermes tools → Computer Use.
You keep your flow. Hermes gets on with its part.
ALT Hermes Computer Use supporting graphic. A large monitor shows a Windows desktop with Microsoft Word on the left and Figma and Outlook on the right. A prominent user cursor remains over Word while a Hermes badge identifies background work. Headline: The cursor stays yours. The graphic highlights cursor, focus and workspace preservation, macOS/Windows/Linux support, local model compatibility, and enabling Computer Use in hermes tools. Real app interfaces and sample content form an illustrative desktop composite.
Cua retweeted
Your AI agent needs its own computer. Not control of your entire laptop.
Cua’s free cua-spaces plugin gives Hermes a separate environment to work in.
- Each run gets a fresh Space by default.
- You can watch live and take over anytime.
- Run it on your hardware including a spare Mac mini or in your cloud.
- The disposable Space is deleted when the run ends.
Hermes skips risky-command prompts inside Spaces the plugin creates. Persistent named Spaces keep those prompts.
That’s the useful shift: fewer interruptions with the agent working separately from your everyday machine.
The plugin is free cloud compute and model usage aren’t included. Shared files and connected accounts still need care.
Give the agent room to work not unrestricted access to everything you own.
hermes-agent.nousresearch.co…
Cua retweeted
目前最强的 desktop use ？
Cua retweeted
A unique benchmark puts user experience in the center: measuring the agent ability to complete task in background without disturbing the user.
Cua retweeted
A new benchmark shows most computer-use AI agents can't work in the background without hijacking your screen. Cua's Driver Bench (built with Scale AI) tested agents on macOS tasks that require staying out of the user's way: Cua passed 58% of runs, OpenAI's Codex computer-use hit 33%, Vercel's agent-desktop 20%, and browser-use just 3%. Foreground-only modes score much higher, showing the gap is specifically about working undisturbed in the background.
Cua retweeted
Running CUA on agentenv!
Cua retweeted
oof, so much worse on linux with that as ai agent dev boxes
year of the linux desktop, it is not. Yet
Cua retweeted
Cua tested AI agents working in the background while a simulated user kept typing on a Mac. Cua's driver passed 58% of the trials. Codex Computer Use passed 33%.
A driver is the part of the agent's software that moves the pointer, clicks and types. The test used Claude Sonnet 5.5 to choose every agent's actions. The tasks, instructions and time limits stayed the same.
A full pass required three things.
A checker confirmed the task was complete.
The agent stayed inside the graphical interface and did not use a shell, which is a way to give commands through text.
The user kept working without interruptions, such as losing keyboard focus, seeing the pointer move or having another window cover theirs.
Each driver ran eight background tasks, such as fixing a schedule, reviewing support cases, reordering a list and handling pop-up messages.
Scale AI separately ran Cua's driver, Vercel Labs' agent-desktop and browser-use on its own AgentEnv system. Cua ran Codex's result. Cua says the two systems are not strictly alike.
The Codex row is OpenAI's driver controlled by Claude, not OpenAI's own agent.
Cua made the winning driver and the test itself. It used these tasks to find and fix problems. A separate hidden task set has not been released.
The top two error ranges touch. Cua says this gap points in its favor but does not prove a lead.
Cua's driver also flashed a pop-up menu over the user's window for about 50 milliseconds. All four drivers still need the real pointer to reveal some controls by moving over them.
The task files, testing code and Scale's trial recordings are public.
