Your AI configuration should be portable, versioned, and yours.
A system of record for the instructions, context, skills, and rules
that decide how your agents behave — broken into atomic assets, versioned
independently, and rendered into whatever format each tool expects.
Files are renderings. Assets are the source of truth.
A CLAUDE.md is not the unit of value — it's an output.
The real substance is smaller: one preference, one constraint, one heuristic, one
fact about how you work. Once those are atomic they can be versioned, encrypted,
recombined, shared selectively, and compiled to any target.
The problem, concretely
Today
The same preference lives in a Claude Code project, a ChatGPT custom
instruction, a Cursor rules file, and an Obsidian note — four syntaxes,
quietly drifting apart. Personal facts sit beside reusable structure, so
sharing means hand-sanitizing first. Nothing answers the basic question:
what is my current setup?
Instead
One canonical set of assets, each marked public, private, or secret.
Compose them into environments. Render those to every target you use.
Change a single rule and see exactly which environments moved, and how.
How it composes
One source of truth, many renderings. Secret values never leave the client in plaintext.
The loop the first version has to prove
Import an existing config — a file, a folder, a repo — without losing the original form.
Review the proposed asset boundaries. The split is AI-assisted; the decision stays yours.
Classify each asset public, private, or secret.
Compose the assets you want into a named environment.
Publish a reusable module, or keep the whole thing private.
Fork someone else's public module — its structure, none of their personal data.
Personalize the values it asks you for.
Render the result to at least two agent-specific formats.
Principles
Privacy is defined at the smallest meaningful asset — not per file, not per repo.
Public reuse never requires sharing private values. Publishing structure is not publishing yourself.
Secrets are encrypted client-side. The service shouldn't hold plaintext it has no need for.
Every AI suggestion is reviewable, reversible, and attributable.
Standalone utility first. It has to be worth using while the public network is still empty.
Open standards over invented ones wherever an existing format will do the job.
What this is not
Not an agent runtime. Not a replacement for Git. Not a prompt
observability platform. Not an optimizer that quietly tunes you without asking.
The first job is the configuration lifecycle — import, version, classify,
compose, render. Marketplace, teams, and governance are all downstream of that
working first.