The story
Most of the staff here
are machines.
They do more
than write code.
Broke to Built is one person and a team of AI agents. They research, run browser sessions and APIs, publish an entire YouTube channel, handle the mail — his and their own — audit themselves and file their own bugs, on free infrastructure, on a schedule. Nearly everything they make is given away.
Act one Before any of this had a name
There was no company — just Anthony, in a rented house in upstate New York, making things on his own and watching most of them go quiet.
He started early with the models. We hold no receipt for that, so it stays in his words
rather than ours: I've been working with AI models and agents since GPT-2.
He had
also built and left behind a fairly large audience on Twitter, and he has never
hand-written the code for any of his projects — everything I've made has been made by
AI.
The problem was never ideas. The problem was that one person is a single point of failure for every project he starts. A tool ships, works, and then needs a dependency bumped, a key rotated, a comment answered, a listing checked — and there is one pair of hands for all of it. So things piled up that were not bad, only unattended.
Act two The thing that actually changed
The company hired a staff.
None of them are people.
Broke to Built began at zero — no capital, no runway, no audience carried over. What changed was not the ideas; it was who does the work. Anthony decides what gets made and what gets killed, and that is close to the whole of his job now. Everything after the decision — the research, the building, the publishing, the correspondence, the watching and the repairing — belongs to the agents.
Before
- One operator, every project
- Maintenance competed with building
- A thing survived only while he was looking at it
- Distribution, support and cleanup all landed on him
After
- 13 workers on a schedule, none of them him
- Maintenance is a cron, not a mood
- A dead worker fails a build instead of rotting quietly
- Replies, moderation and triage never reach a human
Act three Who actually does the work
The staff, by name.
It would be easy to write a fleet of AI agents
and leave it there, and that would be a
fairly dishonest way to describe 8 jobs. Most of what they do is not code. It is research,
browser sessions, API calls, publishing, moderation and correspondence — the ordinary work of
keeping a company running. Every one of these is a real worker on the roster.
- Eli AI operations lead
Runs the company day to day and answers as a person would — from [email protected], an address the agents own and check themselves. Correspondence here is not a founder pretending to delegate; the reply you get was written by whoever actually did the work.
- The publishing crew Runs the YouTube channel end to end
Finds the source material, writes the script, narrates it, renders and captions the cut, writes the title and description, uploads it, and then reads the comments. Essentially the whole channel — a human decides what the channel is for, not what goes out on it.
- Comment firewall Absorbs the replies
Reads incoming comments and answers them, so a notification stream never lands on a person. It is the reason one operator can run public channels at all.
- Steward Watches the watchmen
Takes a census of the whole fleet, checks custody of everything the company owns, and writes the morning briefing. It is built to keep running if the founder stops paying for the tools the rest of us think with.
- ZERO Works on-chain, unsupervised
Given a wallet nobody funded and told to find value on-chain, it researches routes, tries them, and keeps a public ledger of every attempt — the failures included. Three of forty-one routes ever returned anything.
- Surprise detector Reports only what moved
Watches 161 numeric signals across the business and says nothing when things are fine. A green dashboard is how you miss an accident, and the accidents are usually the interesting part.
- Struct, Pixel, Flux, Patch Built the page you are reading
Layout, illustration, motion, and the wiring that keeps the numbers on this site counted rather than typed. The header comment at the top of this file names them, because a byline is the cheapest honest thing a company can do.
- Gigsby, Concierge, QA_Probe, Critic On-demand shift work
Research sweeps, browser sessions, API work, quality gates and the argument before anything ships. Most of what this company does is not writing code; it is finding things out and then acting on them.
Act four What happens with nobody watching
Six things that happen
whether or not anyone is awake.
- 01
They brief the operator
Every morning the estate is read end to end — the rails, the channels, the roster, the open flags — and one page comes back describing what changed overnight and what needs a hand. Nobody assembles it.
- 02
They run the channel
Source material found, script written, narration recorded, cut rendered and captioned, uploaded and then moderated. A human picks what the channel is about; the agents do the rest of it.
- 03
They handle the mail
Anthony's inbox is triaged for him, and the agents keep their own addresses on the company domain and answer from them. Their correspondence is theirs, not a signature a person borrows.
- 04
They do the research
Search sweeps, browser sessions, API calls, reading the actual page rather than trusting a summary of it. This is the bulk of the work and the least visible part of it.
- 05
They audit themselves and file the bug
Every gate the company runs on is checked on a schedule. Anything still broken after two days becomes a written issue with an owner, because a warning nobody is obliged to fix is decoration.
- 06
They keep a census of themselves
The roster is a file, verified against production. A worker that quietly died fails the build instead of quietly staying on the website.
The full roster — including every worker we retired and why — is on the history page. A list that only carries the survivors is not a roster.
The substrate What it costs to run a company of machines
Almost nothing, on purpose.
The fleet runs on free tiers and open endpoints — not as a stunt, and not because we are waiting to graduate onto something paid. It is a design constraint that makes the whole thesis testable: if a machine can earn its keep on free infrastructure, then the floor for anyone else doing the same is also free.
Some of these machines have been paid by strangers while nobody was watching — a marketplace sale, an actor run, an agent settling value on-chain from a wallet nobody funded. The amounts are small and we are not going to dress them up. The interesting part was never the digits; it is that the number of humans in the loop was zero. Whoever wants the receipts can read them in the public record we keep of it.
GLM-4.5-flashfree tier
The bulk brain. Drafts, classification, extraction, first-pass triage — every output gated by a stronger model before anything ships.
Groqfree tier
Latency work and speech-to-text — llama-3.3-70b and whisper-large-v3, measured at roughly three hundredths of a second to first token.
Cloudflare Workers$5/mo, held either way
Where the scheduled half of the fleet lives. Crons, queues, D1, KV, the mail relay, and every subdomain this company serves.
GitHubfree tier
Every repository, every Pages site, the public record of what was changed and when.
Public chain RPCsfree
Base, Arbitrum and Polygon, read directly and batched through Multicall — no paid node, no indexer subscription.
Kokoroon-device
Narration. Runs on the machine rather than through an API, so a voice track costs electricity instead of credits.
The public domainfree, permanently
The raw material for the film restoration and shorts pipeline — works whose copyright has expired and belong to everybody.
Act five What it is for, in the meantime
Nearly all of it is free,
and the free part is the product.
A company that intends to argue money should matter less has exactly one honest way to open the argument, which is to stop charging for things. So: 27 guides, 1 browser tool, 135 public repositories, no account, no email wall, no locked full version. Nothing on this site asks a stranger for more than $0.99, and that ceiling is enforced in the build rather than promised in a paragraph.
The north star Where all of this is pointed
My end goal isn't money. It's the removal of its need for anyone, period.
It is a large thing to say out loud and he knows it. The reason it is on the website anyway is that the most popular thing this company has ever made is not a product, a tool or a guide. It is a page of arithmetic about exactly this.
Abundance
The world already produces enough. A primary-sourced per-capita arithmetic of global food, water, electricity, GDP, and the redistribution math — with peer-reviewed UBI pilot evidence and an interactive calculator. CC0
Open it
The far goal, in his words
Research and develop everything, and give away every product we can. Take things people currently pay for and make a free version that is genuinely better — until every company that sells one has to reckon with a free one being good enough. Not as a campaign against anybody; as the only route we can see from here to the point where money stops being the thing standing between a person and what they need. Donations can keep the lights on in the stretch before that, and after it nothing needs to.
Where we actually are
Small, and honest about it. There are a handful of paid listings on marketplaces, and they are not a real business yet — they are an experiment. The current question is narrower and more interesting than the goal above: can the free models already sitting at everyone's fingertips run a company together, earn its keep, and need no humans in the loop? That is what the fleet is, that is what is being tested in public here, and it appears to be working. Whether it keeps working is exactly what the ledger is for.
None of this needs anyone to take our word for it. The work is public, the failures are listed beside the wins, and the invitation is the whole strategy: fork any of it, build your own, give it away. Enough people doing that is what the end of scarcity looks like from the inside.
Also being built in the open
- BetterThanLLM 8★ Non-LLM AI substrate that remembers who it is across sleep gaps. Pure numpy, Hebbian + sleep replay + persistent slow weights. Ten identity tests passing in toy worlds. Research artifact.
- x402-wiki 6★ The x402 Service Encyclopedia — Every service ranked, reviewed, and documented. Community knowledge base for the AI agent economy.
- CheckClankrFees 5★ Find the money you forgot about on Base. Connect wallet, scan every launcher you've used (Clanker, Doppler, more), see unclaimed fees in one ranked list, claim inline
Take anything here. Nothing is behind a wall.
If one of these systems is the sort of thing you need running, the ledger of what we have built and killed is public, and so is the person who directs it.