node banner.jsa release crossing the line, live

  

regatta

A 220 Labs tool

regatta is an open framework for a business run by agents.
You make the decisions; the agents do the rest.

For a founder or a small team with more ideas than hours. Say an idea into your phone. Agents weigh it against your goals, plan it, build it and ship it. Each piece of work names the numbers it should move: a product metric, a cost, hours of yours. Once it lands, the numbers say whether it did, and the next pick is made on that. That last loop is the next thing we build. Everything you see is shaped to cost you the least time. Only the questions that need you reach you, in one portal: documents per project, cross-referenced, each question with its context and a recommendation. Every answer is kept with your name on it. It runs on Claude Code. Take one module, or all of them.

Public Beta: several small businesses already run on it unattended, for the people who started them.

See the course Read the code

The course

The course A loop with twelve marks at even intervals, sailed clockwise from sparks. The first half is the build, idea to launch: which to sail, the water ahead, the plan, build and launch, runs by itself, to the finish under the blue flag across the course from the start. The second half is the run, launch to the next idea: the public door, measure, decisions, remember, the night, and round again. The lighthouse stands at the centre, on the line between the halves. A sloop sails one slow lap and fires a burst as it crosses the line. BUILD · idea → launchRUN · launch → next ideasparkssextantnavigatorhelmwharfwharf-runsregattahelloworldplumbkeelatlasauroralighthouse The course A loop with twelve marks at even intervals, sailed clockwise from sparks. The first half is the build, idea to launch: which to sail, the water ahead, the plan, build and launch, runs by itself, to the finish under the blue flag across the course from the start. The second half is the run, launch to the next idea: the public door, measure, decisions, remember, the night, and round again. The lighthouse stands at the centre, on the line between the halves. A sloop sails one slow lap and fires a burst as it crosses the line. BUILD · idea → launchRUN · launch → next ideasparkssextantnavigatorhelmwharfwharf-runsregattahelloworldplumbkeelatlasauroralighthouse

Why regatta

  • It keeps going after launch. Most agent frameworks stop at the release. regatta closes the loop. After a launch it reads the numbers, writes down what the project taught, and briefs the next agent with it. The same half runs the shop, from tech reviews to customer support.
  • Attention is all you need. Yours, that is, and the system is built around how it gets used. Only the questions that are really yours reach you: an irreversible choice, money, two good options. You answer in place and the agents carry on. The rest gets built and reported. No chat window sits waiting on you.
  • Everything is on record. Ideas, plans, decisions, releases and lessons are plain text in git, one repo per project, every answer signed. A lesson learnt on a project reaches the next agent on that project the moment it lands. A lesson that holds everywhere reaches every project the next day. Read the history, undo a step, or copy the whole thing.
  • Built like it has to run a big company. Nothing ships without its tests and a review. Nothing goes public without a privacy scan. Each project has its own repos, scope and login. Every decision is written down. When the week's AI budget runs low, new work waits. Google could run on it, but it is tuned for a team of one, and the aim is to take that team from one to zero. Controversial, we know. Large companies already have large integrators to look after them.
  • Take only what you need. Each module is one Claude Code plugin. Install one, or all of them. Knowledge lives in git, so an agent works with what it already knows, a repo and a Markdown file, instead of learning your knowledge graph. No database, no search index, no server to run. The two exceptions are optional: the site bot and the dead-man's switch, each one small web worker on a free plan.

The 220 Labs team supports integrations and pilots, big and small: hello@220labs.dev.

The fine print

  • Small teams. The dependencies are free and open source, and optional until you want the extras. It suits a team that settles most conflicts with a git merge and manages access with what GitHub gives you. spacesheep and the like bring outside people in, but they are no substitute for enterprise access control, and that is not planned. Our aim is a company from one employee to none. One to a thousand is somebody else's race.
  • Claude required. The agents that do the work can run on other vendors' models. The framework around them needs Anthropic's Claude. The core needs Claude Code, git and Python 3.9; GitHub adds pull requests and CI before a merge. A module that needs more says so on its card.
  • macOS first. The human side runs on a Mac for now. An idea comes in through Apple Notes or a Voice Memo, one tap to record. A notice reaches you by iMessage. The night jobs run on launchd. Both this and the Claude line can go as the community grows. For now we use what works for us, and say so in case it is a blocker.

Integrations

Free plans are enough, and so is none: every service here is optional, GitHub too. Pay for a plan and regatta uses what it adds, to the full.

  • spacesheep · Decision pages you answer from your phone, the project board, the Session tab. Optional. We recommend its free plan, and Pro when you run without Orca.
  • Orca · Many agents side by side, each in its own working copy, across your machines. Optional. Free and open source; we recommend it.
  • GitHub · Pull requests, CI before a merge, and the hub between machines. Optional. Any git remote carries the work between machines; pull requests and CI are GitHub's.
  • Cloudflare Workers · The two small servers regatta can run: the site bot and the dead-man's switch. Optional. Its free plan is enough.
  • Telegram · Where you answer your site's visitors, from your phone. Optional, with beacon. Free; its sink needs Resend's free plan.
  • Resend · Mails your reply to a visitor who has left the chat. Optional. Part of beacon's Telegram sink; the free plan covers a small site.
  • CircleCI · One of three places a project's check can run before a merge. Optional.

All integrations: what each brings, by module, its free and paid plans, and regatta without it.

Modules · eleven plugins: take one, or all of them

sparks ideasTurns voice memos and jotted notes into filed ideas

You say an idea into your phone, or type a line into a note, and by the time you sit down it is one note per idea, in plain words, filed where it will be read, committed. It reads the audio and the notes you point it at; nothing else.

It decides nothing about what to build. That is sextant's job, and sparks is sextant's intake: it ships inside the sextant plugin, never as a plugin of its own.

Principles

  • One note per idea, never a transcript.
  • Filed where it will be read, not where it arrived.
  • Redacted before a model reads it, where an inbox opts in.

Today

  • In the sextant plugin: voice memos, written notes and Apple Notes in.
  • Audio transcribed on your machine with whisper.cpp.
  • Redaction is built, a classifier and a regex layer; no inbox has switched it on yet.
  • Each note committed to a knowledge vault or a project's docs.
  • An install script: the launcher, the schedule, the two macOS grants; a chat notice when a note is filed.
sextant strategyPicks which ideas to build, against your goals

Your north star, goals, priorities and budgets in one file. Every week the module reads the numbers, the ideas pool and your ventures against it, turns kept ideas into proposals and the best proposals into candidates with a verdict. A candidate that is safe to build starts on its own, inside the budget you set.

It does not set your goals, and it never launches past a question only you can answer.

Principles

  • The mechanism is public; your strategy stays private.
  • A candidate names its goal, its estimate and its verdict.
  • Nothing unattended past a Reversible: no question.

Today

  • sparks, its ideas intake.
  • north-star.md: your goals, priorities and budgets, parsed and checked.
  • A weekly pass that writes proposals from ideas and numbers into a queue of candidates.
  • A sounding before each launch, checked against the charted shoals.
  • Launches armed after one dry week.
helm planningSplits big work into tasks and keeps them moving

A large idea arrives as one doc. helm breaks it into ships and blocks, turns each block into a task and starts the ready ones through wharf-runs, relays the one question that needs you, and keeps a board that shows what waits on whom. Skippers, the manager agents, run a ship each.

It builds nothing itself and starts no session of its own: wharf-runs does. There is no bare helm command either; Kubernetes owns that name, so the verb is regatta helm.

Principles

  • One skipper per ship; you answer only Reversible: no.
  • A stall it can retry is retried; the rest are reported.
  • The board shows what waits, and on whom.

Today

  • The tick, asks, the board.
  • A skipper per ship, an agent that keeps it moving.
  • A ship starts itself once its course is answered.
  • Groundings counted; a cause seen twice is charted as a shoal.
wharf shippingRecords every step of the agents’ work and ships it after review

One session, one worktree, on its own branch. Every turn that changed a file is committed, with the prompt quoted. A finish curates the branch into clean commits, proves nothing was lost, runs the review gate and launches it through a pull request. main moves no other way, bar the records a repo marks to be written in place.

It needs git and a terminal, and GitHub adds pull requests and CI before a merge. It runs while someone is in the session. Starting work with nobody there is wharf-runs.

Principles

  • main moves only through /finish, by a pull request where GitHub is the remote, bar the records a repo marks in place.
  • Every turn that changed a file is a commit.
  • Git is the only state.

Today

  • The hook, /finish, /todo, the review gate, seven agents.
  • /finish scans the curated commit messages where the repo has a public twin.
  • Needs git; any remote carries work between machines, GitHub adds pull requests.
wharf-runs autopilotStarts agents on work while nobody is at the keyboard

A branch seeded with a brief starts on its own at the time you set, in a fresh session, and ends in a pull request. A chain starts the next step itself. A branch left idle about forty minutes launches itself through /finish, every gate run; a tight week pauses it, bar a branch already deployed, which launches at once.

It needs wharf, and today macOS: the schedule runs on launchd. It does not choose what to build; a brief does.

Principles

  • A run starts fresh, from a brief in git.
  • Every gate stops it; a tight week pauses it.
  • Idle and green launches itself.

Today

  • orca-host, which keeps each Mac's Orca set up and paired.
  • Dispatch, chains, blueprints, the launcher, pickup, three harnesses: Claude Code, Codex, Orca.
  • install.sh --schedule writes the launcher's sweep, every ten minutes.
  • macOS only: launchd runs the schedule.
regatta coreThe shared base every module stands on

Every module needs the same few things: one library, a table that says which model and effort each task gets, a price shown as list and effective, a secrets check, and one command that says what is installed. The core is those things once, so no module carries its own copy.

It runs no business by itself. Install it alone and you have a config directory and a CLI; the work starts with the first module beside it.

Principles

  • One library; no module keeps a copy.
  • A model and its effort are a table row, not a prompt.
  • Hosts, logins and people live in config, never in code.

Today

  • The library, the model gate, cost, secrets, the CLI.
  • regatta <name> runs a tool kept off PATH: regatta pr, regatta helm.
  • install.sh names the clone in the config directory; check says what is missing.
  • Config in one directory: .config/regatta in your home folder, or wherever REGATTA_CONFIG points.
  • A contract check every module passes before the marketplace lists it.
  • Needs Python 3.9, git and Claude Code; macOS first.
helloworld publishingNames a project and publishes it behind a privacy scan

The door from a private repo to a public one. Name the project first: handles, domains and packages probed, then a decision doc. Then publish: the allowlisted files go to a public twin behind a privacy scan of every path, word and commit message, and the scan fails closed. The first push and every merge are yours.

It will not make a repo public for you, and it has no opinion on your code. It gates words, paths and names.

Principles

  • What goes public is an allowlist, never everything-but.
  • A private word never reaches a twin; the gate fails closed.
  • The public history should read like the private one.

Today

  • name-check, /name, publish, and check as the source repo's gate.
  • This repo's own twin is published through it, since 2026-10-19.
  • Replay: each curated commit becomes one public commit, scanned on its own; a failing one folds into the next; one squash only as the fallback.
  • check-message scans a commit message before it reaches main.
  • publish --pr needs wharf for its pre-push guard, and refuses without it.
plumb metricsCounts your time per project and what the agents cost

What the work costs, in dollars and in attention, counted. Your own time per project each week; the agents' usage and cost per project and per week, list and effective; how much of the week is left and a tier hint at the first prompt; how often a session stopped to ask you. Every number comes with the command that made it.

It measures; it does not throttle. The gates that read its numbers live in the modules they guard.

Principles

  • Every price: list and effective.
  • Count what cost attention, not only what cost money.
  • A number has a command that reproduces it.

Today

  • crew-report, your own time per project each week.
  • Usage, cost, the week, interruptions, a status line, a limits card per login.
  • Groundings per week, for helm's board.
  • Benches for decisions and sessions, run by hand.
keel decisionsPuts your questions in one document to answer

Every question a session has for you reaches you in one doc: a numbered list, each item with its options and a recommendation, on its own branch, with a page you can read on your phone. You answer inline. The answers come back as commits, and the session carries on.

It asks only what you must answer: an irreversible choice, two good options, money, or a third party's say. The rest is built and reported. A Reversible: no item is always yours.

Principles

  • One place to answer, each item with a recommendation.
  • Ask only what is irreversible, costly, or a real fork.
  • An answer is a commit, never a chat line.

Today

  • The D-list, keel lint, the page, answers as commits.
  • A reply-ending lint that blocks, once, a reply that asks D-items outside their doc.
  • Reads the other modules through guards, so it works on its own.
atlas memoryWrites down what was learnt, where the next agent looks

What a session learns outlives the session. A fact gets one home: the bulletin every session reads this week, the rules every session loads, a project's docs, the knowledge vault. Promotion moves it up; the records and charts keep the long story. The next session finds it without being told.

Plain files in git: no search index, no embeddings, no search service. It remembers what a session filed, nothing it did not.

Principles

  • One home per fact; the others point at it.
  • Short-lived things expire; rules are dated.
  • A lesson that failed twice becomes a mechanism.

Today

  • The bulletin, /share, /promote, vault commits, the docs map.
  • A vault commit goes through a pull request when wharf is installed; records commit in place.
  • Transcripts archived; records and charts kept by the night job.
aurora reviewReads the week overnight and suggests what to change

Aurora Borealis, the night job. While you sleep it reads the week, memory, bulletin and transcripts, and writes one reviewed diff of what should change in the rules; it updates the charts, digests the week and tends the branches. Monday starts with the reading done.

It changes no rule on its own: the diff waits for you. It works best with atlas, wharf and plumb, and says what is missing without them.

Principles

  • One reviewed diff a week, never a silent edit.
  • The night reads; the morning decides.
  • Every rule change is dated.

Today

  • The weekly loop, the dream, the charts, the digest.
  • /embark, the Monday sitting that applies what you approve.
  • install.sh --schedule writes its nightly job.
lighthouse upkeepKeeps every machine healthy, and answers site visitors

The machines stay in shape without anyone remembering to. Settings are declared in a file and written to each config directory; a health check runs daily and names its fix; logins, trusted projects and windows are set the same way. Drift is reported, never quietly repaired.

It writes only what a declared file says, and it tends one machine at a time; beacon, its support part, is one Cloudflare Worker you deploy to your own account. There is no fleet view.

Principles

  • Declared in a file, applied by a script, drift reported.
  • A finding names its fix.
  • Nothing a person must click twice.

Today

  • Health, declared settings, the login walk, trust, tiling, publish.
  • Four guard hooks: the account, secrets, the design route, the tool path.
  • Shoals, the causes of a stuck run, kept as items.
  • beacon: a site's contact form and its bot, one Cloudflare Worker on your Cloudflare account; it needs Node 22.13+.
  • deadman: a dead-man's switch for the machines that run work unattended, one Cloudflare Worker on the free plan.
  • macOS: the schedule and the window tiling are Mac tools; tiling needs yabai.
  • Needs helloworld as well as the core.

Install

claude plugin marketplace add 220labs/regatta
claude plugin install regatta@regatta
claude plugin install helloworld@regatta

Each plugin needs the core, and the others install the same way, by name; sparks and navigator come inside sextant.