Tell it once.
The work runs itself.
Describe the outcome in a sentence. Melvin works out every step and runs it end to end, across your native Mac apps and the open web, in one job. No connectors to wire up, nothing to script.
Clear the inbox into drafts
Find the prospects, build the list
- Web
- Numbers
Rename and file a whole folder
- Finder
- Preview
Book the table, after you say yes
- Web
Put the dates on the calendar
- Finder
- Preview
- Calendar
Cross the web and the Mac in one run
- Web
- Numbers
- Finder
Draft the replies, send nothing
- Messages
Research it, compare it, table it
- Web
- Numbers
Reply to everything unread. Don’t send any of it.
17 drafts written. Nothing sent.
The numbers
Top tier on the board. And the board is the small part.
- 1Hark HandoffCareerflow.ai97.7
- 2Yutori Navigator n1.5Careerflow.ai97.3
- 3MelvinOur own run, same 300 tasks95.0
- 4Enhans ACT-2Careerflow.ai92.7
- 5Google Computer UseOfficial leaderboard69.0
- 6OpenAI OperatorOfficial leaderboard61.3
Scores as published on the official leaderboard and by Careerflow.ai. Melvin’s 95.0 is our own run of the same 300 tasks, graded the same way. When the independent number is published it replaces this one, whatever it says.
The entire Mac, personalised
The entire Mac, the whole web, and it learns you.
Most agents pick a side: a browser tab, or a script on your files. Melvin runs the entire macOS system and a full cloud browser as one job, so a sentence can start in Mail and finish on a web form without you carrying anything across.
That is the floor. What sets it apart is that it learns you. Melvin builds on your past runs, sees what is on your Mac right now, and carries the job into the future on its own: how you name things, which Sarah is the client, what happens every Tuesday at nine. A generic agent has to be walked through your work. Melvin starts already knowing it.
Saved jobs
Save the job. Run it again.
A good run should not happen only once. Save the task and Melvin will run it again with the same context: the same apps, the same outcome, the same stopping points. Put it on a schedule if the job repeats. Daily, every weekday, or weekly, at a time you pick. You are not rewriting the prompt each morning.
Nothing permanent without you
It stops before the irreversible step.
Every run reads first and changes second, and anything that cannot be undone waits for you: a sent email, a booked table, a deleted file. Everything reversible just runs.
Two permissions to start, Accessibility and Automation. No screen recording and no screenshots: Melvin reads the interface your apps already publish. Web work runs in a browser Melvin operates in the cloud, and every run says which lane it used. Mac work never leaves the Mac.
One sentence. Two lanes. One result.
One run reaches across your Mac and the open web and back again. Every step is checked against the thing itself, never against what Melvin thinks it did.
You saidPull last month's charges off the billing portal, file the receipts, and put every renewal in my calendar.
- 1The billing email 3 attachments seen
- 2Open the portal final URL confirmed
- 3Sign in saved session, no password
- 4Pull 11 invoices 11 files, 4.2 MB
- 5File to Receipts 11 on disk, hashes match
- 6Read every date 11 dates parsed
- 7Reconcile totals agrees to the cent
- 8Update the vendor record read back
- 9Build the sheet saved, 11 rows
- 10Date the renewals 11 events queried back
- 11Send to finance waits for you
4 lane changes in one run. 6 steps on this Mac, 5 in a browser Melvin drives in the cloud. Mac work never leaves the Mac, and every step says which lane it used.
10 checks, none of them self-reported. Files are measured on disk, totals are compared against the portal, events are read back out of Calendar. The last step changes something outside your Mac, so it waits for you.
Questions
Frequently asked questions
What you can hand it, what it will not do on its own, and how to stop a run. If this list misses it, write us.
Ask usBoth, and that is the point. Native work runs on this Mac through the interface each app publishes to macOS: Mail, Finder, Notes, Preview, Calendar, Reminders, Messages, Contacts, Settings, Shortcuts. Web work runs in a full browser Melvin drives on its own, so it can sign in and work a site the way you would. A single job can use both without you handing it over.
Yes. Melvin runs several jobs concurrently, and how many depends on the memory in your Mac. Jobs touching different apps run in parallel; two jobs that both want Mail take turns, so one run never types into the other run’s window.
Yes. Every step is visible as it happens, and you can stop any run on its own without touching the others.
No. Anything permanent stops and asks you first. Drafts stay drafts until you press send. Everything reversible just runs.
No. Melvin reads the accessibility interface, not pixels, so it never takes a screenshot and never asks for Screen Recording. Accessibility and Automation are the two permissions it needs. Screen capture ships turned off and there is no switch in the app that turns it on.
From the lab
From the lab.
Engineering09/03/2026
You will stop prompting. Here is what has to be true first.
An agent doing a thirty step task at 95 percent per step finishes 21 percent of the time. Two measurements put the harness sixteen and eighteen points ahead of the model on identical weights. Verification is the lever, with our own numbers on how far it reaches today.
Privacy & Permissions08/25/2026
What each macOS permission actually lets an app do
A plain English breakdown of every macOS privacy permission an app can request, what each one actually grants, and how to audit what you've already given away.
Engineering08/18/2026
We measured the accessibility tree against screen capture. Both lose.
Same Chrome window, same second. The accessibility API returned the tab strip in 77ms. Screen capture plus OCR returned 17 lines in 509ms. Neither one had the page. Measurements from one Mac, including the failure modes that report success.
Tell it once. The work runs itself.
Describe the outcome. Melvin works out every step and runs it, across your Mac and the web.
Available for Apple Silicon Macs. Windows coming later.