subaud

the guide · the kit · 1 of 3

The tools

You run few of these tools by hand. You say what you want in plain words, and the agent runs the commands. Each entry's ask is something you can paste as it stands, in a terminal or to your agent as its block says; a word in ‹angle marks› is yours to fill in.

A few words come up often. A session is one agent at work in one terminal, for example the agent fixing a bug in your website. pi is the open-source agent program most of our sessions run in; an extension is a package that adds a feature to pi, such as a new tool the agent can call. A skill is a file of instructions an agent reads when a task matches it, for example how to debug.

The subaud tools

The tools in the suite. Each runs on your machine and keeps its state on your disk.

projThe subaud tools

what
Opens each project in its own terminal session, with the agent you choose running in it and the tools' pages beside the terminal. Its list shows every session, which agent is in each, and which are waiting on you; after a reboot it brings the sessions back. It comes two ways: proj in a terminal, and the same list as a page in your browser at 127.0.0.1:51500, which starts when you log in so you can bookmark it. From the page, a session opens in Chrome, with its terminal on the left and the tools' pages as tabs on the right, or in a Ghostty terminal window.
why
One session per piece of work keeps each agent's terminal, notes and pages together, so several pieces of work can run at once without getting mixed up.
ask

proj is one of the few tools we open ourselves, from its page or a terminal, to pick a session or start new work:

Terminal
proj

Your agent can open one too:

Say to your agent
Open a proj session for ‹project› so I can work on ‹what› there.
proj on tackle.tools

creelThe subaud tools

what
A small masked box where you type or paste an API key or password; the value goes straight into a local .env file, the file a project reads its secrets from. When a command needs the key, the agent hands it to that command without reading it.
why
A secret pasted into the chat becomes part of the agent's conversation. With creel the agent asks for the key by name and only ever hears back "added" or "updated".
ask
Say to your agent
I need to give you ‹KEY_NAME›. Ask me for it with creel.
creel on tackle.tools

diazoThe subaud tools

what
The planning page. Your agent lays out the choices in a change as cards, each with its options and a recommendation; you decide on the page, and it ends in a spec and a plan you approve before anything is built.
why
A decision made on the page stays there with your answer, and nothing is built until you have approved the plan. Each step (what the agent understood, the decisions, the spec, the plans) waits for your approval before the next.
ask
Say to your agent
Plan this with diazo: ‹your idea in a sentence›.
diazo.tools

galleyThe subaud tools

what
A document you and your agent revise in rounds. You mark up the draft in your browser, the agent rewrites it, and you read what changed. It works on markdown files and on the words of an HTML page.
why
Writing someone else will read needs more than one pass. galley keeps every round, so you can see each version and what changed between them, and your own edits are never undone by the agent.
ask
Say to your agent
Let's use galley to edit ‹file›.
galley.tools · how it works

musterThe subaud tools

what
A message bus between the agent sessions on one machine. A session can send another a message, or hand it a task and get the result back. It runs over a local socket and never calls a model itself.
why
Without it, you carry text between terminals yourself. With it, one session can tell another that its work has landed, or ask a session with no stake in the code to review it.
ask
Say to your agent
Ask another session to review this branch, and bring me its verdict.
muster.tools · docs

kemptThe subaud tools

what
Machine setup from one readable file. The file lists the tools, settings and files your machine should have; kempt shows what's installed, what's behind, and every change it would make before it makes any.
why
Setting up a new machine like your old one, or getting a new release onto it, becomes one request. Running it twice is safe: if nothing changed, it does nothing.
ask
Terminal
kempt update
kempt.tools · how it works

casebookThe subaud tools · beta

what
A record of what should happen to every repository, pull request, branch and worktree you own. Your agent recommends keep, archive or delete for each, and you decide on casebook's page. It also notes the git activity of every session.
why
Branches, worktrees and pull requests pile up, and it gets hard to tell which are safe to remove. casebook gives each one a decision you made.
ask
Say to your agent
Use casebook to go through my stale branches and recommend what to do with each.
casebook.tools · how it works

siftThe subaud tools · beta

what
Audits the instruction files your agents read: files such as AGENTS.md, which tells every agent in a project how to work, and skills. Your agent recommends a revised version of each file, you decide on a local page, and what you approve becomes one pull request per repository.
why
Instruction files grow, and then agents ignore or trip over them. sift finds what is too big, duplicated, stale or in the wrong place.
ask
Say to your agent
Run a sift round on my agent instruction files.
sift on tackle.tools

cullThe subaud tools · beta

what
Judges your automated tests and removes the ones that don't earn their place. It asks TypeSafe's Jev, a model that answers typed questions, whether each test is worth keeping, cuts the ones it is sure about, and asks you about the rest on a local page.
why
Agents write a lot of tests. After a batch, the suite can feel slow or full of near-duplicates; cull trims it with you deciding the unsure ones.
ask
Say to your agent
Use cull to check the tests I changed on this branch.

mudaThe subaud tools · beta

what
Finds waste in how your code gets delivered: slow or flaky CI (the checks that run on every change), needless waits, and cloud cost. It measures your real runs first, then fixes one finding per pull request, each one approved by you. Muda is Japanese for "waste".
why
You get numbers before changing anything. Every finding is labelled measured or suspected, and every fix is proved on the real pipeline afterwards.
ask
Say to your agent
CI is slow here. Find out why with muda.
muda.tools · docs

hailThe subaud tools · beta

what
A way to reach a running pi session from your phone: see what it's doing, answer the question it's stuck on, and send it back to work. We use it every day, but it isn't public yet.
why
An agent that stops to ask a question waits until you're back at the terminal. hail lets you answer from wherever you are, while the session stays on your machine.
ask
(not public yet)
hail.tools

scratchThe subaud tools

what
A notes pad for each session, kept outside your repository, that you and your agent both write to. It saves as you type.
why
Some things you need in ten minutes, not forever: a link, an error message, the next thing to try. They go on the pad and can never be committed by accident.
ask
Say to your agent
Put a note in my scratch pad: ‹note›.
scratch on tackle.tools

imprintThe subaud tools

what
Guides your agent through publishing npm packages: logging in, publishing, setting a trusted publisher (the one CI workflow allowed to publish a package), deprecating, and confirming a release. It opens each npm approval in your browser.
why
Publishing has several steps that only the package owner can approve. imprint opens each approval for you and checks every step on the registry before calling it done.
ask
Say to your agent
Publish this package to npm with imprint.
imprint on tackle.tools

channelsThe subaud tools

what
A pi extension, published as channels.tools, that lets the other programs on your machine reach a running session. You connect channel servers, small local programs that push events in, and an event wakes the session when it is idle or waits until its current turn ends. muster and galley both speak it.
why
Mail from another session, a review round landing, a build finishing: without channels, each has to be noticed and pasted in by hand. With it, the session hears about them itself.
ask
Say to your agent
Which channel servers is this session connected to?
channels.tools · how it works · source

billetThe subaud tools

what
Project workspaces in your own AWS account. Each workspace is a machine with your repos cloned, the suite installed and kept current by kempt, and your agent, which you open from a browser or your phone after signing in with a passkey.
why
Your work isn't tied to one laptop. A diazo plan moves from your Mac to a workspace and back, through GitHub, and the people you add to a workspace each get their own home there.
ask
Say to your agent
Set up billet in my AWS account.
billet.tools