Entity OS: Run Your Companies on Files You Own

There is a particular feeling that visits the owner of a few small companies at around ten o’clock on a Sunday night. A question surfaces — did the LLC actually reimburse me for that repair, or did I pay it personally? When is the nonprofit’s annual filing due? Could the holding company cover that expense this quarter? — and with it comes the honest answer, which is that you don’t know, and finding out means seven logins, a folder of PDFs named things like scan_0043.pdf, a shoebox, and forty-five minutes you don’t have. So the question goes back under the surface, and it takes a little of your sleep with it.

I know that feeling well, because I run a small family portfolio — a holding company, a couple of operating companies, a nonprofit we started, and the household books besides — and for years my records were exactly that scattered. What I eventually built to replace the shoebox is the thing this article announces: Entity OS, an open specification for running a small portfolio of entities on plain text files maintained by an AI clerk. It’s on my GitHub now, free, under licenses that let you take all of it.

But the specification is downstream of the reason, so let me start with the reason.

What I actually wanted was peace of mind

I didn’t want software. I wanted a set of feelings I didn’t have:

I wanted to know where everything stands, at any hour, in seconds. Not approximately — exactly, with the receipts one click away. The Sunday-night question should have a Sunday-night answer.

I wanted my books to confess their own errors. Merchants solved this five hundred years ago: Luca Pacioli’s Venetian traders entered every transaction twice so the ledger would check itself, because their ships were at sea and they could not afford to be wrong. I wanted that — records that announce when they don’t balance, rather than waiting for April to embarrass me.

I wanted my family to be able to pick it up without me. If I’m unavailable — traveling, sick, or worse — my wife should not inherit a scavenger hunt. The system keeps a one-page, plain-language guide at the front: what this is, where each company’s records live, who to call. The bus factor of the family’s affairs should not be one.

I wanted nothing held hostage. Every dashboard I subscribe to is a tenancy: it can reprice, get acquired, or lock me out. If every vendor I use vanished tomorrow, I wanted my books, my minutes, my filings history — all of it — still sitting on my own disk, readable by any human being with a text editor.

I wanted a history that remembers. Every change to every file, timestamped, forever. What did the books say on March 3rd? That should be a question the records answer, not a reconstruction project.

That’s the product. Everything else is how.

What it is, in plain words

Every entity — each LLC, the nonprofit, even the household — gets the same simple anatomy, and all of it is plain files:

One drawer that never changes. Bank statements, signed documents, prior-year returns go into a folder that is never edited. Originals stay original.

A set of readable pages the AI keeps current. An assistant reads those sources and maintains a small wiki for each entity — what it owns, what it owes, what happened and when — cross-linked and written in ordinary prose. Ask a question, get a cited answer drawn from documents, not from memory or vibes.

Books that check themselves. The ledger is plain text in a format a validator can prove balanced — Pacioli’s self-confessing ledger, enforced by a program. From it come the same statements a bookkeeper would hand you.

A corporate binder that’s actually current. Minutes, resolutions, the filings log, the record of what the owners put in and took out — kept as tidy files on a fixed pattern, so “did we ever document that?” is a thirty-second lookup instead of an archaeology dig the week before a deadline.

And one rule above all the parts: the clerk never gets the last word. The AI files, drafts, and cross-links; deterministic checks decide what’s true; and anything it can’t categorize with confidence goes into a named pile — mine is literally called Ask My Accountant — and waits for a human. It never guesses. A system earns trust not by being impressive but by admitting what it doesn’t know, every single time.

What peace of mind turned out to look like

Some moments from the far side, because this is what I’d want to know before building it:

The Sunday-night question now takes under a minute, and the answer comes with sources. That alone repaid the build.

Tax season stopped being a reckoning. My CPA gets clean statements and organized records instead of a shoebox, which means our conversations start at the interesting questions instead of at “what is this deposit?” — better advice, less billable excavation. And when I sit down with my attorney about the structure of things, I walk in with the entity’s whole documented life in hand. Organized records don’t replace professionals; they make you a dramatically better client for them, and the professionals’ time goes where it’s actually valuable. There’s a well-known bill many attorneys charge businesses that arrive with years of undocumented decisions — the “entity clean-up” engagement, reconstructing minutes and records that should have existed all along. Records kept current as life happens are how you never receive it.

The nonprofit’s filing deadline lives in a log a script can sweep, not in my head. The difference between those two locations is measured in cortisol.

And the whole estate is versioned like software, because it deserves the same audit trail software gets. Every change, every date, forever, on my disk.

Take it

Entity OS is published at github.com/PLACEHOLDER/entity-os — Apache 2.0 for the code and schemas, CC BY 4.0 for the prose. The repository ships with a complete worked example: Juniper Holdings LLC, a fully synthetic little company — wiki, corporate binder, a ledger that validates to the penny — with every name, number, and date invented. Copy it, rename it, and you have entity number one by the weekend.

It has two siblings extracted from the same working system: the delegate-downwards ratchet, for anyone running recurring AI work who wants it to get cheaper on evidence, and the presuppositional LLM wiki, for anyone building a knowledge base that should know what it believes. Each stands alone.

If you’d rather have it built for you

Everything above is free and always will be. But building a system is its own project, and some readers are past the do-it-yourself season of life. Installing this discipline for other people is part of my work now: my agency, Clineo, runs websites, operations, and automation for independent healthcare practices on exactly the philosophy in this article — automation that never guesses, checks that decide, a human gating everything that matters. If you run a practice, the Automation ROI Calculator takes about two minutes and will tell you what your repetitive work is actually costing you; and if the number annoys you enough, book a call — the conversation is free.

If you own two LLCs and a shoebox of PDFs — if you know the Sunday-night feeling — you are the person I wrote this for. The cure turned out to be files.

I am an operator, not an attorney. This is not legal advice.