A plugin for Claude

An engineer in the conversation.

Your team builds its own apps with Claude. Steadyship is the engineer working beside them, always: it asks what they don't know to ask, judges what they can't judge, and brings in real engineering the moment an app starts to matter. Not before.

Open source. Stays on your computer. Keeps working if you remove it.

What an engineer brings

Claude builds what it is asked to build. These are the three things nobody in the building was there to add.

Asks what you don't know to ask

Before a change is built, it has read your business notes, your other apps and your spreadsheets. So it asks the one or two questions that would be expensive to get wrong, recommends an answer, and builds once you decide.

You say "Make the job board track hours." It asks: against the whole job, or each step? And isn't that already in Hours.csv on the shared drive?

Judges what you can't judge

A straight Yes, No or Not yet on whether the team can rely on an app, with the evidence. Security in plain words: who can see this, where what people type ends up, whether a password is sitting in a file.

It never calls an app secure. It says what it checked, and labels what it could not.

Brings engineering when it's needed, not before

A prototype stays a prototype. The day someone relies on it, it gets one master for its data, a backup, a test, a named owner and its decisions written down, each step with one plain line on why. The apps stop drifting apart.

A one-person tool gets no ceremony. The day the workshop runs on it, it gets looked after like software.

What it does

Each one is a plain sentence in Claude. No dashboard, no account, nothing to learn. Everything it finds goes into files you keep.

01

Asks before it builds

The one or two questions that would be expensive to get wrong, with a recommendation. You decide, it builds, and it writes down why.

02

Gives a straight verdict

Yes, No or Not yet on whether the team can rely on an app, with the evidence. A second reviewer checks the first.

03

Finds the same thing kept twice

Reads every app and every spreadsheet and shows where a price, a job or a name is in two places and already disagrees.

04

Checks the ground first

Can it reach GitHub, read and write the shared drive, open your app? One minute, PASS or FAIL, written down.

05

A path, not a rebuild

Prototype, owned tool, proper app, platform. It says which stage an app is at and moves it one step at a time, with you deciding at each.

06

Apps that stay in step

A shared glossary with a human owner. Every app reading the same spreadsheets instead of its own copy.

07

A second opinion from another vendor

A reviewer from a different AI vendor, so it does not share the builder's blind spots. No terminal needed.

08

Everything stays yours

Ordinary files in a private folder the business owns. Remove it tomorrow and everything it wrote is still there.

How it works

Three sentences on day one. After that, ask for changes the way you already do.

  1. 1

    “Set up Steadyship for Harbour Boatworks”

    Answer four questions. About two minutes.

  2. 2

    “Check this computer”

    Three checks, one minute, recorded.

  3. 3

    “Review the job board”

    A verdict the owner can act on.

See it answer

A real reply from the demo business, a boat builder with six apps its staff built with Claude. The short version is what you read. The full picture, every row that differs, goes into a file you keep.

Try it on the demo business
What apps do we have and where is the same information kept twice?

The three things that matter most

  1. The Quote Builder is using old prices. Price List 2026.csv has Bimini top $3,450 and Trailer $6,250. The app has $3,200 and $5,800, so a quote with both is $700 low before GST.
  2. The job list is typed into several places, and they disagree. Jobs.csv has 8 jobs. Job Board has 8, Delivery Checklist 4 and Shop Timer 3. J-1039 is “Sea trial” in Jobs.csv but “Paint” on the board.
  3. If a laptop died, most of the work would be lost. Five apps keep what people enter only in that browser on that computer.

What I'd do first

Update the Quote Builder's two prices, and make it read Price List 2026.csv. It is the one customers see numbers from.

The full picture is in docs/assessment/SUMMARY.md. Say ‘show me everything’ and I'll put it here.

From prototype to something the business runs on

Every app is at one of four stages. Steadyship says which, what would move it on, and does the moving one step at a time with a review at each step. Nobody has to choose a framework, a database or a host. It makes the call, gives the reason, and you can defer.

  1. Stage 1

    Prototype

    One file, one person's tool.

    Moves on when someone else starts using it.

  2. Stage 2

    Owned tool

    Shared, backed-up data with one writer. A test. A named owner.

    Moves on when two apps touch the same information.

  3. Stage 3

    Proper app

    A real framework, shared parts, data checked on the way in.

    Moves on when the business depends on it daily.

  4. Stage 4

    Platform

    Several apps on one database, kept in step.

    Moves on when they must stay consistent for the business to run.

Always there, never in the way. Exploring? Prototype as you do today; it asks at most once and never blocks. Keeping it? It gets a name, an owner and a first verdict. Relying on it? The stages run, with you deciding at each. Short answers, with more if you ask.

How each step works

Who it's for

A business where a few people who are not developers build their own tools with Claude, and those tools are starting to matter. A job board someone relies on. A quote builder that disagrees with the price list. Two apps that both keep the customer list.

One person does a ten-minute setup. The owner decides what the business relies on. Everyone else keeps building, with an engineer in the conversation.

What it is not

  • Not a place to host apps, and not a database.
  • Not a replacement for Claude. It sits in the same chat.
  • Not a certificate that an app is secure. It tells you what it checked and what it could not.
  • Not in the way. It stays quiet until something matters.

Ten minutes to a first verdict.

Windows or Mac, the Claude desktop app or Claude Code in a terminal. One page for each.