_jects — the data layer that builds itself

Your AI runs on files. Give them somewhere better than a Git repo.

Skills, agents, prompts, runbooks, playbooks — the markdown your team’s AI actually runs on. Today it lives in repositories, where access is all or nothing and every consumer has to clone. Here it lives in a tree where every branch carries its own permission, and any agent reads it in place.

Acme Restaurants 4 folders · 9 files

CLAUDE.md Internal today

Every row carries its own access. No fork, no second repository.

Every audience gets its own branch. Your staff see one set, your produce supplier another, your guests a public page. One tree with different doors — not four copies quietly drifting apart.
Agents read it where it lives. No clone, no sync, no build step. Claude and any other MCP client opens the files directly, with exactly the access the person behind it has.
Publish a branch to the world, or don’t. Public reads arrive on a connection that structurally cannot see anything private — not a filter somebody has to remember to write.
Take it back and mean it. Revoke access and it is gone. Delete a file and the bytes leave storage, not just the link to them.

You talk. It builds.

It works like a colleague who knows your data inside out. Say what you want to build and it tells you what you will need, raises what you missed, settles the plan with you — and then builds it.

Ask before you build. It tells you what you already have, what fits, and which choices get expensive later.
Nothing to design first. No structure to plan, nothing to rebuild later, no storage decision — not on day one, and not ever.
Nobody builds alone. Your coworkers’ agents and skills work on the same store, so the second person to arrive inherits structure instead of negotiating an integration.

you "we need to build a reservation system

for our restaurants"

claude Great. You will need three shapes:

location, table, and booking. Sound right?

you "yes"

claude Have you considered walk-ins? They take a

table without ever making a booking.

you "good idea. can we hold two tables back?"

claude Absolutely. That is a rule on the shape,

not code. Building it now.

claude ✓ 4 shapes · a hold rule · access set

Call it money. It behaves like money.

Nine data shapes come with their own machinery. Say a field is a location and you get distance, what is nearby, and maps. Say it is money and you get exact math, currencies, debits and credits — and never a rounding error. You implemented none of it.

The behavior arrives with the shape. Distance and nearby search, schedules that survive the clocks changing, share links for files. Nothing to wire up, nothing to keep in step.
Validation you never wrote. Describe a field once and every write is checked against it — from your app, from your API, from an agent. One description, enforced everywhere, with no validation code to write and none to keep in sync.
And all of it is an API. Query by shape, publish a subtree, and the OpenAPI document writes itself — public or private, your call.

locationsdistance · what is nearby · maps moneyexact math · currencies · debits and credits · never a rounding error filesuploads · share links · deletion that reaches the bytes schedules“every second Tuesday” · survives the clocks changing amountsnumbers that carry their unit · 5 kg, never just 5

nine in all · a tenth is a decision, not a ticket

Enterprise-grade security, from the first record.

Not a tier you upgrade into, and not a discipline your team has to remember. Permission is enforced inside the database itself, so a mistake in application code cannot reach another customer’s data.

Nobody can forget it. Isolation lives in the database, not in application code. A missed line returns nothing — never somebody else’s records.
Audit-ready by default. The database writes its own change history, with who and when, so it cannot be skipped. Evidence for SOC 2 or HIPAA is a query, not a project.
As fine-grained as you need. Share one branch with one person, a department, or an outside partner, using roles you define. The same model runs a private tool and a public product — no security tier to buy.

privatethe owner, and whoever they name internaleveryone in your organization publicanybody at all, once you publish

and on any branch of the tree

granta person, a group, a department, a partner roleverbs you bundle and name yourself branchone rule governs everything beneath it, until you split it

enforced by the database, not by the code that asked

AI made life simpler. Now it does the same for your data storage.

Your AI runs on files. Give them somewhere better than a Git repo.

You talk. It builds.

Call it money. It behaves like money.

Enterprise-grade security, from the first record.

Start with one record.