macOS app
Put every agent in one room, and watch them work.
Wonop runs Claude, Codex and Grok side by side on your Mac — in real browsers, terminals and documents you can see, and step into whenever you want. Four agents working. One window to watch.
Apple Silicon · $99 during the beta · pay once, no subscription
The 11pm problem
Four things running. No idea where any of them are.
Starting agents got easy. Everyone can hand out four jobs at once now.
Keeping hold of them didn't. Four conversations open, tabbing between them to see whether anything has finished, never quite able to focus on your own work because something might need you.
Then there's the reading. Every agent writes confidently, whether it did the job or quietly got it wrong — so you can't skim any of it. Nothing tells you which part deserves your attention, so you give it to all of it. That costs more than the agents saved.
And there's no good moment to stop. Nothing says the work is finished, so you keep poking at it, or you walk away and hope.
How it works
One window. Everything happening inside it.
Wonop opens real applications on your Mac — a browser, a terminal, documents, project folders — and lets your agents work in them beside you. Not a description of the browsing. The browsing.
Each workspace binds its own windows, terminals and files. Two projects open at once never contaminate each other — and neither do two agents in the same project.
Nothing to read
You don't read what happened. You watch it happen.
An agent that writes you a report hasn't saved you the work. It's moved it — now you're reading two thousand words to find the one line that went wrong.
Wonop doesn't narrate. Every action is a row with its state on it: running, done, waiting for you, refused. You look at the colour, not the prose. The two that need you are the two that look different.
Failures are loud
Most tools are built to look reliable. That's the wrong goal — a system that never looks wrong makes you worse at noticing when it is. So when something here fails, it's the loudest thing on screen.
| Run | Call | Agent | State | Elapsed |
| 04a1 | wm.browser.capture | claude | done | 2.41s |
| 04a2 | wm.term.run | codex | running | 11m 06s |
| 04a3 | wm.fs.patch | codex | refused | 0.38s |
| 04a4 | wm.pdf.extract | claude | done | 6.02s |
| 04a5 | wm.window.arrange | you | done | 0.09s |
Refused: the agent didn't have permission for this. It was told why. So were you.
Knowing when to stop
It tells you when it's actually finished.
Work is held properly, not in a chat thread. A task pauses when it needs a decision from you and is still waiting tomorrow. It survives a restart, halfway through. And finished means finished with the receipts attached — change something underneath and it stops counting as done.
- Handed over
You asked for it
Given to an agent with its own workspace and its own windows.
- In flight
Waiting for you
It hit a decision it will not make alone. It is waiting, not guessing.
- Done
Finished, with proof
Completion arrives with evidence attached, not just a claim.
- Stale
Done, until it wasn't
Something underneath changed. It stops counting as finished.
On your machine
It can only touch what you hand it.
Wonop runs on your Mac. Your files stay on your Mac. Anything an agent can reach beyond that is something you granted, one capability at a time — which sites, which commands, which folders. These aren't promises in a privacy policy. They're enforced at the moment the action is attempted.
- command
- Which calls the package may make at all.
- host
- Which hosts it may reach over the network.
- pty
- Whether it may open a terminal, and which one.
- argv
- Which arguments it may pass to a binary.
No new subscription
It uses the AI you already pay for.
Sign in with the ChatGPT, Claude or Grok account you already have. Wonop runs them; it doesn't resell them, and it never adds a bill per message. Which also means nobody is metering your work — only possible because Wonop never sits between you and the model in the first place.
The plan you already have
Your existing subscription is the plan. No keys to buy, no credits to top up.
More than one at a time
Put two providers on the same job and compare them yourself, instead of being handed one answer.
Nothing metered
Wonop takes no cut per message, because it never handles the message.
Integrations
Sorted by what they can reach.
Every integration is a package, and a package can only touch what its manifest declares. So this list isn't sorted by logo — it's sorted by reach, and the right-hand column is what actually gets enforced.
Reach the network3
- LinearList, filter, open, create, edit and re-status Linear issues.api.linear.app + key
- GitHub IssuesBrowse, open, comment on and close GitHub issues.api.github.com + token
- Epidemic SoundSearch and play music from your Epidemic Sound library.epidemicsound.com + key
Drive a terminal3
- Claude AgentRun the Claude coding agent in a workbench window, on your own subscription.pty
- Codex AgentRun the OpenAI Codex CLI in a workbench window, on your own subscription.pty
- Grok AgentRun the Grok CLI in a workbench window, on your own login.pty
Never leave this machine6
- ShellRun a command and read what it printed.pty
- Git LogRecent commits and status for a repository.git log, git status
- MarkdownRead Markdown files rendered, with an outline and links you can follow.local files
- Design BenchA board of design screens and versions, with annotations an agent can act on.local files
- ArrangeTidy the windows on this workspace.windows
- Local WorkThe built-in provider for custom, local work items.nothing
Asked for, not built yet1
- Post BridgeScheduling and cross-posting to social accounts from a workspace.on the list
Twelve of these ship with the app. Anything missing is a small file away, and it shows up in this table with its own reach spelled out, same as the rest.
Pricing
Pay once. No subscription.
One purchase, and a year of updates. After that the app keeps working — you only pay again if you want another year of new things. Running several agents usually means renting a higher tier every month, on top of the AI subscriptions you already have. Wonop is $99, once, and adds nothing to what those accounts already cost.
Two hundred dollars off while Wonop is in beta. At 1.0 it goes to $299.
- Every window, package and command in the app
- As many agents at once as your Mac will run — no per-agent tier
- One year of updates, then it is yours to keep using
- The AI accounts you already have — no model billing on top
- All your own Macs, no seats, no per-message cost
Why there's no usage tier
Charging per run would mean sitting between you and your AI account, which would mean seeing the work. Wonop is built not to. The pricing follows the architecture.
Why one price and not three
There's one product. Tiering it would mean holding parts of it back from the people paying least, and there's nothing in here worth holding back.
Where it actually is
Early, and specific about it.
This is version 0.1.0. Here's what isn't done, so you hear it now rather than on day three.
- macOS
- Apple Silicon only. No Intel build, no Windows, no Linux, and no date for one.
- Editor
- The built-in editor is a placeholder. People keep their own editor open beside Wonop.
- Packages
- You can install packages you or your team wrote. There is no signed marketplace yet, and we are not going to call it one.
- Under the hood
- Wonop shows you the actual commands agents run. That is the point, but it does mean the app looks technical in places.
Questions
The ones people actually ask.
Trial terms are still being decided — ask us before you buy if it matters.
$99 while Wonop is in beta
Stop wondering what they're up to.
One download, one price, no account to create. It runs on the Mac in front of you, with the AI you already pay for.
Apple Silicon · 0.1.0 · pay once, no subscription