The one record of what every AI did in every app a person works in.
Kept by the one party that sells neither the model nor the app.
Every model company now ships an agent that acts on a computer, and every app company puts its own AI inside its own app. Each keeps a record that stops at the edge of its own models or its own apps. A record of what every AI did in every app can't belong to any one of them. Being neutral across your rivals is the one thing a rival structurally can't do.
yarnnn is that record, and the product people use to make it. It looks like the AI they already use — a chat, a plan, a question when it needs one — on purpose. What is different is where it reaches: inside their own Word, Excel and PowerPoint, their Figma, their browser signed in as them, through each app's own door, on whichever engine they pick. Every change is signed by whoever made it, human or not, into one history the person owns.
Reach is why a person chooses it. The two neutralities are the position no model maker or app maker can take. The record is the proof of both — and, with tenure, the moat.
Stage & traction
Beta — in daily use, by us first
The workspace, the desktop app for Mac and Windows, the Chrome extension and the MCP connector all ship today. We run our own company on it, and we publish where it falls short.
Over 500 decisions, in the open
Every architectural decision is numbered and recorded, with the test that holds it. Who changed what is enforced at the one place a change is written, not claimed in copy.
$500K pre-seed · $5M pre-money
Raised to take the beta to its first paying customers: make the work on a person's own files and programs dependable, learn what they pay for, and grow from one person to the team around them. The full deck and terms on request.
The case here is structural: a position the platforms can't take for reasons of incentive, not capability. It is not a market-sizing argument.
What the wave brings, and what it doesn't
The platforms now market memory, schedules and agents that act on a desktop. Those capabilities arrive in waves, and we borrow the shell rather than compete on it. What doesn't arrive with the wave is one agent inside every app as the person, on any engine, and the neutral record of what every engine did there.
A platform's own agent
The company that builds the agent also keeps its record, and reaches work three ways: a computer of its own, the person's screen, or its own app. Memory you can't read. Changes with no author you can check. And a record that stops at the edge of that company's models or its apps.
Capability parity is real and arrives in waves. Structure doesn't wash out.
Native reach, and the neutral record
One agent inside every app the person works in, through each app's own door, as them, on any engine. One history the person owns, where every change names who made it. Fix something once, and everything after starts from the fix.
yarnnn — the asking, the doing, and the record
Model-agnostic neutrality by construction; theirs by impossibility.
Memory startups have the opposite problem: they remember, but the work is done somewhere else. And the workspaces that added AI afterwards ask you to move the work into their editor. yarnnn goes to the file you already have.
Enforced in code, not claimed in copy
Four properties give the position its shape. Each is a rule the code holds, with a test that fails when it stops being true.
One way in for every change
Nothing reaches the record except as a signed change that names who made it, what it changed and what came before. A person, an agent and a connected AI all pass the same door.
Nothing acts on its own
An agent works when a person asks, or on repeating work a person set up. It can't widen its own access or raise its own budget.
The person keeps the machine
On a person's computer and in their browser, consent is given once, on that machine, and stays there. Every step after it is listed under the person's name.
Each workspace is its own
The record is the customer's. Nothing is learned across workspaces, nothing trains a model, and all of it exports as a plain git repository.
What's live
Work done where it is, on any engine
A conversation where a person asks. An agent that changes their Word, Excel and PowerPoint files where they are, works in a Mac's programs, and acts on websites in their own Chrome. And a signed record of every change, wherever it was made.
Desktop app and extension
Mac, Windows and Chrome. Each is paired to one person, works in the background, and lists every step it takes.
Work that repeats
A person says what should stay current, and when. It runs on schedule, checked against the instructions they gave, and every run is listed.
Model-agnostic, MCP-native
Engines from five providers inside, picked per conversation. Outside, any AI that speaks MCP reads and writes the same workspace under its own name. The record grows with the ecosystem, not against it.
Leave with everything
The whole workspace exports as a git repository with its history. People can leave, which is why they stay.
Investment thesis
Engines commoditize on a quarterly cycle, and so does the shell: every model company ships a chat, a plan, and an agent on a desktop. We borrow that shell on purpose, so there is nothing to learn. What we build is reach — inside the apps, as the person, on any engine — and what we keep is the record of it.
Neither a model maker nor an app maker can hold that position. One audits only its own models and reaches work by a sandbox or the screen; the other goes inside only its own apps, on its own model. A party that sells neither can go inside every app on any engine and keep the one record of what all of them did. Git made distributed collaboration trustable because history was the trust layer. This is that ledger, for work done by people and AI together, inside the tools the work already lives in.
Reach is the differentiator. The two neutralities are the position. The record is the proof — and, with tenure, the moat. The wave that washes out most AI-app moats strengthens this one: every new frontier engine, and every new app a person works in, deepens the record while staying swappable.
Motion
The buyer is a psychographic, not an occupation: someone who already works with AI every day, and is still carrying its output, by hand, to the files and people it was for.
One person first
A Personal plan: one monthly price with usage included, bought by the person who uses it. No procurement, no pilot, no admin.
Then the team
A Team workspace is an organization's own: an admin, a seat for each person, one bill. Priced per person and bought from the pricing page.
Two revenue legs
The plan, and usage beyond it. Usage is the model work an agent runs, so it grows with how much work a person hands over.
The product does the selling: it is useful to one person on the first day, with nobody to approve the purchase. The record is what a second person, and then an organization, joins.
Founder
Kevin Kim — Solo Founder & CEO
Korean-born, US-based. A decade of work spanning enterprise systems, cross-border operations, and context architecture — from deploying CRM for Japan Tobacco in post-military Myanmar to building GTM systems for cross-border sales teams.
Shipped the entire product solo: the web application (Next.js, FastAPI, Supabase), the desktop app for Mac and Windows, the Chrome extension, the MCP server, and the signed write path under all of it — documented across more than five hundred Architecture Decision Records and run on its own operations before raising a dollar.
Let's talk.
If you're investing in AI infrastructure, the accountability layer, or the future of agent-first work — I'd love to share the deck and walk through the architecture.
kvkthecreator@gmail.com