The command-line tool

Mirepoix has a command-line tool, mirepoix, that talks to the same agent API Agent access describes. It’s for connecting your own AI model to a kitchen, or for scripting — reading your recipes, your Menu, your shopping list, or your Pantry, and changing them, from a terminal instead of a browser.

Before you start

You need a kitchen with agent access turned on by its owner, and Ruby 3.2 or newer on your computer. Check what you have:

ruby -v

On a Mac, the Ruby that comes with the system is too old. Install a newer one with Homebrew:

brew install ruby

Homebrew’s Ruby doesn’t put itself on your PATH automatically, so add it yourself. Add this line to ~/.zshrc, open a new terminal, and check ruby -v again:

export PATH="$(brew --prefix ruby)/bin:$PATH"

Install it

In a terminal:

mkdir -p ~/.local/bin
curl -fsSLo ~/.local/bin/mirepoix https://mirepoix.recipes/cli/mirepoix
chmod +x ~/.local/bin/mirepoix
~/.local/bin/mirepoix version

The last line should print a short line of JSON naming the tool’s own version, such as {"cli":"1.0.0","supported_api_major":1,"ruby":"3.4.0"}. If you see a page of syntax error lines, or a message saying mirepoix needs Ruby 3.2 or newer, the ruby it found is too old — check ruby -v and make sure a newer one comes first on your PATH.

Add it to your PATH, so you can type mirepoix instead of the full path. Add this to your shell’s startup file (~/.bashrc, ~/.zshrc, or similar), then open a new terminal:

export PATH="$HOME/.local/bin:$PATH"

Verify the download, if you want to. Compare its checksum with the one Mirepoix’s server publishes alongside it:

curl -fsSL https://mirepoix.recipes/cli/mirepoix.sha256
shasum -a 256 ~/.local/bin/mirepoix

(On Linux, use sha256sum ~/.local/bin/mirepoix in place of shasum -a 256.) The first word of each line should match. This catches a corrupted or truncated download; it doesn’t prove the server itself is trustworthy.

Connect it to a kitchen

Two ways:

Approve it, the way you’d approve a new device. Run mirepoix auth login. It prints an eight-character code and a page address. Open that address in a browser where you’re already signed in to Mirepoix, type the code, and approve the request — choosing which kitchen it can use and whether it’s read-only or can also make changes, the same as connecting a command-line tool by approving it. The tool saves its token once you approve.
Paste in a token you created on your Account page. Create one under Agents and devices, as described in Create a token, then run mirepoix auth add-token and paste it when asked.

Either way, the token is saved for next time — you don’t repeat this for every command.

Naming the connection. Without a name of your own, mirepoix auth login defaults to “mirepoix CLI on” followed by your computer’s name. That’s the name you check on the approval page, the name that shows on your Account page afterward, and — once the tool adds or changes a recipe — the name that shows in that recipe’s history. Give it a name you’ll recognize instead with --label:

mirepoix auth login --label "Work laptop"

Tell your AI model what the tool can do

Run mirepoix briefing. It prints how the tool expects to be driven: finding operations, previewing a write before committing it and when it’s safe to skip that, reading exit codes, and the few rules worth knowing before giving it to an AI model to drive. Most agent setups work best by handing it this output once, up front, rather than making the model guess.

Everyday commands

mirepoix ops lists every operation the API offers, with a one-line description of each.
mirepoix describe <op> shows one operation’s full detail: what input it takes, what it returns, and what can go wrong.
mirepoix call <op> runs one. A read runs right away. A write — anything that changes your kitchen — needs --preview to see what would happen without it happening, or --commit to make it happen. Sending the exact input a preview returned, with --commit, is how you turn a preview into a real change.

For example, reading your Menu:

mirepoix call menu.get

Previewing before adding something to your shopping list:

mirepoix call groceries.add --input '{"items": ["olive oil"]}' --preview

Add --human to mirepoix call to read the result as text instead of JSON — handy when you’re trying a command before handing it to a model.

Check the connection, or disconnect

mirepoix auth status says which kitchen the saved token is for, and whether it can currently make changes.
mirepoix auth logout revokes the token on Mirepoix’s servers, then forgets the saved copy. If Mirepoix can’t confirm the revoke went through — a network problem, say — the tool keeps the token and tells you to revoke it on your Account page yourself; add --forget-anyway to forget the saved copy regardless.

Updating

mirepoix self-update

It downloads the current version from Mirepoix’s own server, checks it against its published checksum, and replaces itself — only when a newer version is available, and only from mirepoix.recipes. Add --check to see whether an update is available without installing it.

What leaves the app

Whatever the tool reads with your token goes wherever you send it — to an AI model you’re driving with it, for example — under that destination’s own terms, the same as any other app or tool connected with a token. See What leaves the app for what that can include, and the privacy page’s Agent access (beta) section for what Mirepoix keeps about a token versus what the tool itself keeps on your own computer.