Enter the closed beta.

A local-first thinking-continuity tool that runs entirely on your own machine. One invite admits one person. Enter your code to pick up your build.

Invite code

Your code is checked against a stored hash — it is never saved, logged, or echoed in plaintext. This page collects nothing else: no account, no email, no analytics.

Step 1 · Download

Your beta package

One zip, built from source. It contains the app, a launcher, and a start script — nothing else, and none of anyone else's data.

Step 2 · Start

One command

You’ll need macOS or Linux and Python 3.11+ (macOS ships 3.9 — install a newer one from python.org or brew install python if needed). Then, from the folder where the zip downloaded:

unzip cognitive-atoms-beta.zip && cd cognitive-atoms-beta && pip install -e ".[dev]" && bash scripts/beta_user_start.sh.example

The script asks for your invite code, stores it hashed on your machine, starts the app with live AI off, and opens your browser. Already using port 8765? Prefix the last part with CA_BETA_PORT=8801 and use that port.

At the entry gate that opens, enter this same invite code once more. That in-app step is the one that admits you — one code, one person, on your machine. Unlocking this page doesn’t use it up.

Step 3 · The guide

Closed beta setup guide

What this beta is

A local-first thinking-continuity tool that runs entirely on your own machine. Your data never leaves your computer.
A controlled, invite-only test of one path: enter your invite → quick onboarding → capture or import some of your own thinking → see a first insight that is cited back to your own words → enter the product.
An honesty test. Where the tool doesn’t have enough to say, it should say so — never make something up.

What this beta is not

Not a public or hosted product. There’s no account, no cloud, no sign-up.
Not live AI. The AI provider is off for this beta — cognition is computed locally from your own captured words, not generated by a model.
Not the marketing website, and not the mobile app. This is the local desktop app only.
Not multi-user. It admits exactly one person — you — on your machine.

Setup, step by step (≈ 5 minutes)

Download the zip (step 1 above) and open a Terminal in the folder where it landed.
Run the one command (step 2 above). It unzips, installs, and starts the beta. A virtualenv is recommended if that’s part of your usual setup.
Paste your invite code when the script asks, then press Enter. Your code is stored hashed, never as plain text. Your browser opens to the entry gate.
Enter the same invite code at the entry gate. That admits you (one code, one person). If it says “already used” or “not recognised”, reply to your invite email.
Onboard and test. Answer the short prompts, then capture or import some of your own thinking — paste a thought, or import a conversation. The first insight you see will be cited to your own words, or honestly say “still forming” if there isn’t enough yet.

Live AI is off — by design

Chat shows an honest “not connected” state rather than a generated reply, and every insight you see is derived from your own captured atoms, never invented. Even if you happen to have AI keys in your environment, the app will not make a live call — that’s enforced in code.

You start with a blank slate

Your fresh runtime root begins empty. You won’t see anyone else’s notes, profile, or “cognition” — only what you capture. Until you add your own thinking, the cognition screens will honestly read as empty / “forming.”

What to test (and what to tell us)

Just use it like it’s yours and note anything that stands out:

Does onboarding make sense? Where did you hesitate?
Capture from a paste and from a conversation import — did it land faithfully?
The first cited insight: did it feel accurate and genuinely yours? Did the citation point at something you actually said?
Browse Home (continuity) and Explore (cited signals) — anything that feels stale, generic, or made-up? That’s the most important thing to flag.
Try Chat — confirm it’s honestly “off”, not faking answers.
Check the privacy/export/delete language — is it honest about what it does?

The one rule we care about most: if anything feels fabricated, stale, or not-actually-yours, tell us immediately — that’s the bug we’re hunting.

What data not to use

Use your real thinking — that’s what makes the test meaningful — but during this beta, don’t paste secrets (passwords, API keys), regulated/PII data, or third-party-confidential / client-confidential material. It stays on your machine, but it’s a beta; keep it to your own non-sensitive thinking.

How to stop / delete everything

You’re always in control:

Quit the app — Ctrl-C in the terminal running it.
Erase your beta data entirely — delete your runtime folder: rm -rf "$HOME/.cognitive-atoms-beta". That removes everything this beta stored. (You can export first from the app’s Profile/export option if you want to keep a copy.)

How to send feedback

Reply to the email that carried your invite. Anything is useful — especially screenshots of anything that felt false, stale, confusing, or genuinely valuable. Rough notes are perfect; you don’t need to write them up.