The age of software factories
What will your factory ship?
Describe the product. A team of agents plans, builds, tests, and reviews it — inside a budget it cannot exceed. A reviewer signs it off, or it stops.
- Runs on Cloudflare's network
- Every build sandboxed
- Hard budget ceilings
- Reviewed before anything ships
- Changes merged
- 20
- Model spend per merged change
- $0.26
- Changes the reviewer refused
- 10 of 26
Each one a pull request, on a repository you own, that you can still read.
Everything it spent divided by the changes that shipped — not the runs that finished. Container time is metered separately and is not in here.
Sent back for repair or handed to a person. It is not a rubber stamp.
Read from the factory's own database, 17 August 2026.
- 01
Describe it
Plain language is fine. The architect turns your brief into a backlog you can read — and asks when it isn't sure.
- 02
Watch it get built
Engineers work in isolated sandboxes while you watch the plan, the tests, and the spend — live, not a spinner.
- 03
Approve it, and it ships
An independent reviewer checks every change. Approved work merges and deploys to a URL you can open.
Why Shipyard
Prototyping tools make demos.
Factories make products.
v0, Bolt, and Lovable are great for a Friday-afternoon prototype. Shipyard is built for the application your business will still be running in five years.
Progress you can trust
A factory you can see into.
Every run streams its plan, its checks, and its spend. When something isn't working, Shipyard says so plainly — an unfinished task is a short report, never a surprise bill.
And "done" is decided by your project's own tests — not by a model's opinion of itself.
✓ planned — 4 tasks from your brief, ceiling $5.00/day
✓ 1/4 invoice model + migration $0.31 · 9 tests passing
✓ 2/4 QuickBooks export $0.56 · review approved
● 3/4 PDF templates running checks
○ merge → deploy shipyard-app-invoice-app-38be4d.workers.dev
Security
Bold claims, honestly made.
Letting software write software is a real decision. Here is exactly what protects you — and the one thing we refuse to promise.
Isolated by construction
Every build runs in a fresh sandbox on Cloudflare's infrastructure — created for the run, destroyed after it. Your systems are never in the blast radius.
Your credentials stay out
The sandbox never holds your tokens. Repository access passes through an authenticated proxy outside the box — code inside can't leak a credential it never had.
A budget that cannot be exceeded
Spend ceilings are enforced in the payment path itself, per task and per run. Not a policy someone follows. A precondition the code checks.
Nothing ships that the reviewer rejected
Every change is a pull request with an audit trail. A reviewer that never saw the engineer's reasoning has to approve it before it can land, work it refuses comes to you instead, and anything touching the factory's own rules waits for a person whatever else is true. Approved work does merge on its own — the pull request is your record of it, and reverting is one click.
Communication
Written like a person, because we check.
No confetti. No "I hope this email finds you well." When Shipyard writes to you — a progress note, a pull request description, a question — it's brief, specific, and readable in one pass.
If a message reads like it was churned out by a machine, we treat that as a bug and fix it like one.
Morning —
The QuickBooks export is done and tested: 14 files changed, all checks green.
One thing worth your attention: I renamed tax_rate to
tax_rate_bp (basis points) to avoid float rounding on totals.
PR #12 explains the tradeoff.
— Shipyard · $1.14 spent of your $2.50 ceiling
Questions
Asked and answered.
What does it cost?
Today, nothing. Signing up is free, there is no card, and no invitation to wait for — we are watching real factories run before pricing them. What you can already see is the arithmetic we will price from: the ledger above is $0.26 of model spend per change that shipped. When there is a price it will be per change, and you will meet it before it costs you anything.
Who owns the code?
You do. Work lands in your repository, under your license. If you leave Shipyard tomorrow, you keep everything — including the deployed app.
What happens when the factory gets stuck?
It stops and says so, plainly, with what it tried and what it spent. Unfinished work is a short report, never a surprise bill.
Which stacks do you support?
TypeScript and Node.js applications first, and more as we learn which ones people actually ask for. Tell us what you run — [email protected] reaches a person.
Is my code used to train models?
No.
Put our team on it.
Describe the product. Watch a team plan it, build it, review it, and ship it — inside a budget it cannot exceed.