Coming soon — early access
Run a team of AI agents. Without the terminal.
Early access, no spam. One email when we launch.
You're on the list.
We'll email you the moment early access opens. No spam, unsubscribe anytime.
Preview
Nothing to show yet. The crew puts its work here — the page it changes, the shop it runs.
Show the latest workor type an address above
How it works
From plain language to a working diff
- 1
Describe the task
No commands to memorize. Type what you need done, the way you’d explain it to a colleague.
- 2
Agents split the work
Shellcrew hands pieces of the job to separate agents — one reads, two build, one checks.
- 3
Watch it happen
Follow the crew in sentences, and watch your own page change in the panel beside them.
- 4
Review and ship
Approve what needs a key, review the diff, or commit it yourself. The crew never pushes without you.
Why this doesn’t exist yet
AI coding agents are powerful. The tools around them are not for you.
Claude Code, Codex and similar agents can plan, write and run real work — but today you reach them through VS Code, Cursor or a bare terminal. Those tools assume you already know what a shell is, what a git branch is, what a process log means. If you don’t code, the most capable AI agents on the planet are simply out of reach.
Built for developers
Every interface around coding agents today assumes command-line fluency you were never asked to have.
One agent at a time
Most tools give you a single chat with a single agent. Running a coordinated crew means juggling windows by hand.
You have to invent the workflow
The tools hand you agents and leave the hard part to you: who reads the project, who writes, who checks the work. Most people never get past the defaults.
What we don’t build
The product is defined by what it leaves out
Three things every tool in this category ships, deliberately missing here. These, not a feature list, are what separate Shellcrew from what you can already download today.
a terminal tab
Other tools put a chat on one side and a terminal on the other. We take that slot for what the crew knows about your project. Watching four shells at once isn’t a feature for you, it’s stress.
code on screen
No diffs, no stack traces in the default view. If you can’t judge them, showing them by default only manufactures anxiety. They stay one click away for the day you want them.
a model picker
No model name in the status row, no list of nineteen to choose between. The cheap agent reads your project, the expensive one writes. That’s a workflow decision, not yours to make.
What we’re building
A crew of agents, working on your files, in the open
Here’s what early access will unlock — we’re building it now, with the waitlist.
A team of agents, not one chat window
Spin up several Claude Code or Codex agents at once, each on its own task, coordinated from one place.
A real terminal underneath
Every agent runs in an actual shell — nothing simulated. What it does is what actually happens on your machine. You never have to open it.
Watch the work, not the output
Follow the crew in plain language: what it read, what it changed, what it checked. The raw log stays one click away when you want it.
Your own API keys, in a local safe
Keys stay on your machine in an encrypted vault. Shellcrew never sees or stores them on our servers.
Your files, your git history
Agents work directly on your existing folders and repos. Nothing gets copied to a third-party cloud.
Project memory that persists
Shellcrew remembers what a project is, what’s been tried and what’s decided — across sessions, not just one chat.
You decide what the crew can touch
Set how far the agents may go — read only, propose changes, or make them. Nothing lands in your project until you allow it.
Why a crew
Why it’s a crew and not one agent
Each member gets its own copy of your project, works alone in it, and the finished work is folded back in. Nobody overwrites anybody — which is also the only honest answer to “why a crew”: the roles differ because the copies and the permissions differ.
Your project — untouched while they work
scout
copy · read only
Reads everything, changes nothing. Cheap model, runs first.
builder · form
copy 1
Rewrites the email check. Can’t see copy 2.
builder · emails
copy 2
Fixes the confirmation that never sent. Can’t see copy 1.
Folded back into one change — then the checker reads it, then you decide
The word for this underneath is a git worktree, and you never see it. On screen it’s “its own copy” — the same fact, in a language you already have.
What the crew remembers
It remembers what didn’t work
Every assistant promises to remember what you said. None of them promise to remember what failed. If you can’t read the code, the crew proposing the same doomed fix for the third time is indistinguishable from the tool being broken — so the notes keep the dead ends too.
Notes · My shop
How it fits together
Signups go through the form, then an email confirms them. Both had to change for the Safari fix.
You decided
Keep the old signups; don’t ask existing customers to confirm again.
Dead end — don’t try again
Validating emails with a stricter pattern. It rejected addresses with a dot in the name, which is most of them.
Still true
The shop runs on your own machine. Nothing is published until you publish it.
Written by the crew as it works, in the same plain language as the run. You can read and edit all of it, and it carries over to the next task — not just the next message.
While you’re away
Long jobs mean you’ll leave the screen
A real task runs long enough that you’ll go and do something else. So the crew has two ways to work and you pick one per task — not a slider, not a percentage.
Autopilot
The captain merges finished work, hands out the next task, and stops only for a decision that’s actually yours. It stops by itself after twelve steps — no runaway loop, no bill you never agreed to.
Ask first
Nothing moves without your click. Finished work waits in a queue until you open it.
Either way
Nothing is saved until the checks pass. Every task registers its checks before the crew starts, so a change that fails them never reaches your project — it goes back to the crew with the reason attached.
Work that comes back for a second pass is normal, not a failure. We’d rather count it into the time we quote than pretend the first try always lands.
The terminal is still down there. You just never have to open it.
What you hand it
Everything you can hand the crew
Every job that fits in one sentence — poured onto one pile. Grab one and throw it.
Who it’s for
For people who have real work, not a CS degree
Founders
Turn a product idea into working code without hiring an engineer first.
Marketers
Automate the busywork — reports, scrapers, campaign tools — without filing a ticket with dev.
Designers
Wire up a working prototype yourself, then hand off something real instead of a mockup.
Why this exists
This is my own setup, packaged.
I’ve spent months running Claude Code, Codex and Cursor on my own projects. Most of that time didn’t go into writing code — it went into working out how to get the agents to work together without me watching: which one reads the codebase, which one writes, which one checks the work before it lands.
That setup is the product. Shellcrew is the same workflow, with the terminal taken off the front.
FAQ
Straight answers
Does Shellcrew exist yet?
No — not yet. We’re building it for the people who join this waitlist. Signing up shapes what gets built first.
How much will it cost?
Shellcrew’s own price isn’t decided yet — waitlist members hear it first. What is decided: you bring your own AI account and it bills you directly. We never resell it and never mark it up, so the running cost is whatever Anthropic or OpenAI charges you, with nothing added on top.
Do I need to know how to code?
No. You describe what you want in plain language. The agents write and run the code, in a terminal you can watch but never have to type into.
Where do my files and API keys go?
Nowhere but your machine. Files stay in your folders, keys stay in a local encrypted vault. Shellcrew is a control layer, not a cloud that hosts your project.
Which agents will it support?
Claude Code and Codex first, since they’re what most people are already using. More agent backends can follow based on demand.
Will it run on Windows?
Mac first, Windows after. Shipping both at once would hold the first release back, and the Mac build is where the work is happening now.
Want in when it’s ready?
Join the waitlist — one email, when early access opens.
Early access, no spam. One email when we launch.
You're on the list.
We'll email you the moment early access opens. No spam, unsubscribe anytime.