Thanks to visit codestin.com
Credit goes to keeldiscovery.com

Say what you think is true. Keel finds out.

Three sentences — the problem, what you would build, what people would pay. Keel turns each one into lines with a number in it, writes the questions, builds the interview, reads every answer with your own AI, and tells you where your number landed: smaller, bigger, or less often than you thought. It is free for a single founder: your own AI does the thinking, so there is nothing to charge you for.

One real occasion, never an opinion. A guess is shown and never counted.

You type three sentences. Keel does the rest: puts a number in each one, writes the lines and the questions, builds the interview, reads every answer, and tells you where your number landed.

See the whole thing, start to brief →

The founder’s price claim sitting in the composer, your AI’s own question above it, and the two claims already approved, quoted back.

Example data, from real runs: Mulch Run, a landscaping supply idea in Dallas, and Countly, twelve restaurant managers asked.

Your AI’s question — will they pay, and how — and under it the founder’s own sentence, typed in the composer.
You say what you expect. Your AI asks what it needs to make it measurable.

See the whole thing, start to brief →

Start in a few minutes

  1. 1

    Install the plugin in Claude Code

    These run in Claude Code, in your terminal. The Claude app on its own does not install it; if claude is not a command yet, install it once: npm install -g @anthropic-ai/claude-code

    claude plugin marketplace add keeldiscovery/keel-marketplace
    claude plugin install keel@keel
  2. 2

    Type claude to open it, then say “keel connect”

    It asks first whether it may run the plugin's script; say yes. It prints a short code and the link that approves it.

  3. 3

    Approve it on the page it opens.

Python 3.9 or newer is the only thing you need that you might not have.

Free. Your own AI does the thinking, on your machine — Claude Code, Copilot or Codex. You bring five to twelve people who would know.

Nothing to enter a card for.

Works where you already work

One skill, one runtime, and the same two words in whichever AI you have open. Supported means every one of four tests is green; in testing means it installs and runs while one is still being measured.

Claude Code
Supported
CLI 2.1.268 · macOS, Windows
tested 11 Sep 2026
Copilot CLI
Supported
CLI 1.0.83 · macOS, Windows
tested 13 Sep 2026
Codex
Supported
CLI 0.154.0 · macOS, Windows
tested 13 Sep 2026
Python 3.9 or newer is the only thing you need that you might not have.

Also through Spec Kit, as the keel extension — the same skill, on whichever of the AIs above your project uses; the install line is under Get started.

How it works

From a hunch to a brief

You board unsure.

  1. 1

    Keel frames it.

    Problem, solution and price go from vague to a sentence with a number in it. Keel nudges until each has one.

  2. 2

    Keel breaks it down.

    Each sentence becomes lines that can be measured, with the question that measures each, and which ones are deal-breakers.

  3. 3

    Keel builds the interview.

    One short page per person, about one real occasion, never your number. You read it once, approve, and send the links yourself.

  4. 4

    They answer.

    The people you invite open the link and answer in a few minutes.

  5. 5

    Keel reads the answers.

    Your own AI reads what came back on your machine, shows the guessers hollow and never counts them, and updates each line.

  6. 6

    Keel tells you where you landed.

    Each card comes back holding up, not holding up, or split, with the drift — and Keel writes the brief for whoever builds it.

You dock knowing where your numbers landed.

Ten minutes, once

Fit the keel to your own ship.

Connect Claude Code, Copilot or Codex on your own machine. Nothing you learn is locked into a model of ours.

A bare hull.

  1. 1

    Your AI.

    Claude Code, GitHub Copilot or Codex on your machine.

  2. 2

    Fetch the keel.

    Two commands from the marketplace.

  3. 3

    Say “keel connect”.

    The runtime starts on your machine and prints a code.

  4. 4

    Sign in with Google.

    Approve the device on the page it opens.

  5. 5

    Linked.

    A line now runs between your AI and Keel; every reading travels down it and back. Keel keeps your project, the lines, the answers and the record.

  6. 6

    Your machine does the thinking.

    Your AI does the thinking, on your own account. Keel picks the model for each step — a small, fast one for the readings — and never calls an AI itself. It is free for a single founder.

Under way, on your own engine.

See one real idea, end to end →

Keel is in beta. Claude Code, Copilot and Codex today. Spec Kit when it is published.

Start in a few minutes

  1. 1

    Install the plugin in Claude Code

    These run in Claude Code, in your terminal. The Claude app on its own does not install it; if claude is not a command yet, install it once: npm install -g @anthropic-ai/claude-code

    claude plugin marketplace add keeldiscovery/keel-marketplace
    claude plugin install keel@keel
  2. 2

    Type claude to open it, then say “keel connect”

    It asks first whether it may run the plugin's script; say yes. It prints a short code and the link that approves it.

  3. 3

    Approve it on the page it opens.

Python 3.9 or newer is the only thing you need that you might not have.

Free. Your own AI does the thinking, on your machine — Claude Code, Copilot or Codex. You bring five to twelve people who would know.

Nothing to enter a card for.

Is Keel for you?

For you if

  • You have an idea, and five to twelve people who would know.
  • You use Claude Code, Copilot or Codex.
  • You would rather find out now than in six months.

Not for you if

  • You want a survey tool.
  • You want a market-size number.
  • You want to be told you are ready.

Keel will never

  • Score you, rate you, or congratulate you.
  • Email anyone on your behalf.
  • Change a line once someone has been asked about it.
  • Call anything until five people have described a real case.