Portable across tools. Totally private. Totally free.
Stop reteaching models what you need them to know.
Pull it in from the tools you already use, or write it down. Everything stays on your machine until you choose to keep it.
Drop your CLAUDE.md, AGENTS.md, .cursorrules — or any context file or notes you've built yourself.
Ask your AI tools to give you all the context they have on you, or copy the memory they keep in their settings (ChatGPT, Claude, and the rest) — then paste it here. Or use this space to add anything else you want every model to understand about you and what you need from them.
Two ways, depending on you. Both are yours and private — we store nothing.
One click builds a private repo in your GitHub, with your context arranged into a structure Manifest designs for it — we call it a Context OS. In plain terms: it gives every AI tool you use the same context, tells each one to load only what it needs for what you're doing right now, and keeps itself current as you work, handing you updates to approve. More below.
No GitHub needed — one unified text context you carry by hand. It even suggests its own updates; you just save them yourself.
Tell your tool to use the manifest.
Paste this into any tool that can reach your repo — a coding tool opened in it, or a chat tool connected to your GitHub.
First time on a new repo, your tool sets up your context with you — a couple of questions, then your approval. After that, this just loads it. (Some tools need a nudge through the steps.)
Paste or attach your manifest.md (from Step 2). It already tells the model to use your context — nothing else to add.
When you wrap up, have the model save anything durable, so your context grows with you.
Your Context OS already knows how to update itself; this just kicks it off, pointed at your exact repo.
It proposes the changes for your approval.
You hold your context. Nothing is ever sent to us.
Paste this so the model drafts the additions for you to review and save.
Two things, plainly. Your manifest is your portable context — who you are, what you're working on, how you like to work, where you're headed. The name comes from the cargo manifest a ship carries port to port: one list of everything aboard. Yours travels tool to tool.
Your Context OS is that manifest plus the layer that makes it work everywhere — load levels so a model pulls only what a task needs, adapters so it behaves the same in Claude Code, Cursor, or ChatGPT, a privacy model, and an update loop that keeps it current. All plain markdown, human-readable, entirely yours.
Manifest builds it. You keep it.
You don't start from a blank page — you start from what your tools already know.
Pull your memory out of the chat tools you use. ChatGPT and Perplexity expose it in settings; with Claude or Gemini, ask "summarize everything you know about me and how I work" and copy the answer. Drop in any context files you've already written — CLAUDE.md, AGENTS.md, .cursorrules. Add anything else in one open field: notes, your résumé, whatever you'd want every model to know.
However it arrives, it lands exactly as you wrote it. Manifest captures your context; it never rewrites it.
Two options, depending on how far you want to take it.
Text only. Your whole context as a single manifest.md — download it or copy it. It opens with a short "how to use this," then your context. Paste it into a chat or attach it to a project and it works the moment the model reads it. No GitHub, no account.
Connected (recommended). Authorize Manifest once and it writes the full Context OS into a private repo in your own GitHub — folders and all, in a click. Same context as the file, but living instead of static: it loads by task, behaves per-tool, and updates as you work.
Your imported files land in the repo whole, in a sources/ folder — no keyword parser guessing where things go.
The structure gets built the first time you point an AI tool at the repo. Your Context OS ships a SETUP.md that tells the tool to read everything in sources/ and sort it into the real files: who you are, what you're working on, your tech stack, a folder per project, private overlays for anything sensitive. It drafts, flags its guesses, asks you a few gap questions, and waits for your approval. Nothing is invented.
It's a short collaboration, not a single button — a couple of questions, one approval. After that, your context just loads. Manifest captures; your AI tool organizes and maintains.
Chat tools. Kept the file? Paste or attach manifest.md — it already tells the model what to do with it. Connected GitHub? Paste the prompt that names your repo (a connected tool can see all your repos, so "my repo" won't do — the address is what picks the right one).
Coding tools. Claude Code, Cursor, and Codex read the CLAUDE.md / AGENTS.md in whatever repo you open. If your Context OS is that repo, it's automatic. If not, drop a one-line pointer to it in a file the tool already reads, and it loads every session.
Stuffing your entire context into every prompt is the bloat this exists to kill. So MANIFEST.md sorts your context into levels, and a tool pulls only the smallest bundle the task actually needs:
Less spend per prompt, and the sensitive material stays out of the room until you put it there on purpose.
At the end of a session, copy the end-of-session prompt and the model drafts any durable additions for you to review. Or just tell any model "update my manifest." In a connected repo, your tool can do it with your approval.
One honest limit: a true "remind me before I close" depends on the tool. Manifest bakes a nudge into your context so models prompt you within a session, but a hard cross-tool reminder isn't something a portable file can enforce.
Over time, many models touch your context — Claude, ChatGPT, a coding agent, all of them. Each sees only a slice of you, so no version is right just because it came last. Context OS treats every tool as a partial witness, not an author.
So it converges instead of piling up contradictions. While a fact is still unsettled it carries its provenance — observed, reported, inferred, confirmed; once it's confirmed, it becomes plain current-state language and the scaffolding drops away. Settled facts stay settled. Your direct correction always wins, immediately. The same useful context emerges no matter which tool got there first.
Everything you build stays in your browser, on your machine. No server stores your context; we never see it. Connect GitHub and your token lives only in your browser, scoped to the one repo, writing to a private repo in your own account. The only things that ever leave your machine are what you choose to send a model and what you write to your own repo.
We save nothing and gather nothing. Privacy is in the build, not a policy.
* What a model retains on its own end varies — but nothing in your manifest is shared unless you choose to share it.
Save your Context OS to your own private repo, in one click. The connection is between your browser and GitHub — nothing is sent to us.
Create a fine-grained token (Repository access → your repo; Permissions › Repository → Contents: Read and write), then paste your repo and token.