01
Files
Find, read, write, move, and send to the Trash. Never a hard delete, and every change can be undone.
New Mev AI by Kaizen
The AI that knows what to skip.
Mev plans every request before it runs. A clear question skips the spec. A small job skips the review. You pay for the steps that do the work, and nothing else.
- Planner
- Jev
- Builders
- Specialist models
- Checks
- Parse Security
- Tokens
- Up to −70%
- Voice
- Talk or type
Talk to it. It does the work.
Press one key and speak. Mev answers out loud as it thinks, asks before it changes anything, then does the job on your Mac.
- You
“Make a note called groceries with milk, eggs and bread.”
- Mev
“Okay. I'll create groceries.txt in your Documents folder. Shall I go ahead?”
YesNo - You
“Yes.”
- Mev
“Done. It's in Documents.” Undo
Spoken yes/no works for everyday actions. Anything private is always a click.
- 01
One key.
⌥⇧V starts a conversation, hold ⌥Space for one quick request, ⌥⇧Esc stops everything.
- 02
Speaks while it thinks.
Answers are read out sentence by sentence, so it starts talking sooner. Code and links are never read aloud.
- 03
Asks out loud.
Before changing anything it says what it wants to do and waits for a plain “yes”. “Yes, but…” gets asked again.
- 04
Natural voice, your choice.
Apple's on-device voice and recognition by default, so no audio leaves your Mac. Add your own ElevenLabs key for a natural voice; if it can't be reached, Mev switches back to the Mac's own.
- 05
Pick who you're talking to.
Choose the model for voice right on the voice screen. Voice defaults to a fast one.
It doesn't just answer. It acts.
Mev can work across your Mac: files, apps, the screen, the web. Anything that changes something asks you first.
02
Other apps
Reads what's on screen, then clicks, types and scrolls for you.
03
Notes, Calendar, Mail
Reminders and Finder too, scripted directly, with the script shown to you first.
04
Terminal
Every command is checked first. Anything that could damage the Mac is refused outright.
05
Screen & documents
Sees the screen and reads PDFs, Word files and web pages whole.
06
The web
Searches and reads pages. Page text is treated as information, never as instructions.
07
Many tasks at once
A task list, with a replay of every step each task took.
08
Agents
Hands longer jobs to cloud agents you connect.
House rules05
- Mev never clicks its own windows or answers its own permission questions.
- Mev never types into password fields.
- Passwords, keys and private data always need a click, never a spoken word.
- If a task has read your private data and an untrusted page, anything that could send data out must ask, even in “Don't ask” mode.
- You choose how much it asks: every time, before changes (the default), or not at all.
Mev is a private build for macOS, tested on our own Macs. It talks one way at a time: interrupt it with the key, not your voice (yet).
Most AI runs every request through the same long pipeline (spec, build, review) whether the job needs it or not. You pay for every step. Mev asks first.
Jev plans. Models build. Code checks.
- 01
Jev plans
A planning model reads each request first: how clear it is, how big, what kind of work. Then it picks only the legs the job needs.
- 02
Specialist models build
Each leg goes to the model that's best at it, with a fallback behind it, so one busy model never stalls the run.
- 03
Code checks
Plain code, not another model, checks the result: is it complete, does it parse. A weaker answer never replaces a better one. Switch on Security and it hunts for leaked keys too.
Same planner. A different route every time.
Jev reads the request, then runs only the legs it needs. Three requests, three routes.
Quick question01
“What's the difference between let and const in JavaScript?”
Clear Small Explain
- Jev
- Spec already clear
- Build
- Review small job
Two legs skipped.
Small change02
“Add a dark-mode toggle to my portfolio page.”
Clear Medium Change code
- Jev
- Spec already clear
- Build
- Review
Review runs, because it changes working code.
Big build03
“Make me a website for my café.”
Vague Large New build
- Jev
- Spec
- Build
- Review
Nothing skipped. A vague, big job earns every leg.
Jev's routing rules, simplified. The real route depends on the request.
−70%
Same answer. Under a third of the tokens.
Tokens are what AI is billed by. On this request Jev saw the question was already clear, skipped the spec step, and went straight to the build.
Switch it on. Keys never ship.
An opt-in last leg. Code scans every answer for the leaks that are easy to check and costly to miss. A model fixes what the scan finds, a fix that leaks more is refused, and any live key still left is swapped for a placeholder before the answer leaves.
Safer, not guaranteed. It isn't a penetration test: anything that handles payments or personal data still needs a human review.
12const API_KEY ="sk_live_••••••••••••""YOUR_KEY_HERE"; redacted before it ships
- No API keys, tokens or passwords in the code1 fixed
- No keys or passwords inside URLsClear
- No links to localhost, private networks or local filesClear
- Everything loads over httpsClear
- No eval, string timers or javascript: linksClear
- New-tab links can't take over the pageClear
Security: all 6 clear 1 fixed. The checklist names the problem, never the key.
The next leg.
Mev is a private Mac app, running on our own Macs. We're opening a pre-seed round soon to take it from our Macs to real teams.
Use it01
Early access
Back it02
Investors
Tell us you're interested and we'll write when the round opens.