Privacy Architecture
We read your screen,
not the inside of your apps.
The floor
Four things we never do
The dangerous way to "read chat history" is to reach inside the chat app — attach to its process, decrypt its local database, impersonate its client. That can read what you aren't even looking at, and it leaves a detectable, ban-worthy trace on the platform. OKEight does none of it — these are capabilities that simply don't exist in our code.
Never attach to the process or read its memory
We don't attach to your chat app or read its running memory. We can't — and don't try to — reach the history you aren't looking at.
Never read or decrypt the local database
Your encrypted chat archives on disk are untouched. There is no code path that reads or decrypts your chat app's files.
Never connect to the chat platform's servers
To the platform, you look like a normal user of the official client: no fake client, no anomalous traffic, no signal that could get an account banned.
Never touch your account credentials
We don't log into your account and never need your password or tokens. Your identity stays strictly between you and your chat app.
How the data flows
One read, mostly on your device
From "glance at a contact and their tone" to a finished result, the data passes through four steps — the first two never leave your Mac.
Screen capture
Via a macOS system API (ScreenCaptureKit), capturing only the window shown on your screen. The OS does the capture; your chat app receives no signal that it happened.
Text recognition
Apple's built-in Vision framework turns the picture into text, on your Mac. No network, ~0.1–0.3s, nothing leaves your device.
AI reads attribution & tone
The extracted text is assembled into a scrubbed payload and passed to an AI model to judge who said what, and the tone. Where it goes is your choice — see below.
Stored locally
The relationship and tone notes it derives live in files on your own Mac. They are not uploaded to OKEight's servers.
Delete before spend: a contact you delete is caught before any AI call happens — never re-analyzed, never billed, never stored.
AI processing · your choice
Where step 3 goes is up to you
Only the "who said what, and in what tone" step needs an AI model. You have full control over where that data goes.
Relayed through OKEight's secure proxy
A scrubbed text payload is forwarded to an AI model through our secure proxy — no key setup, works out of the box.
Straight to your own provider, bypassing us
Add your own AI provider key and the data goes directly to your own model provider, bypassing OKEight entirely — we don't even relay it. For the highest privacy, choose this.
Permissions & verifiability
Granting access is between you and your OS.
Our promises don't require your blind trust.
Screen Recording — used to capture the chat window shown on your screen. macOS prompts you clearly; you can revoke it any time in System Settings.
Accessibility — used to bring the chat window forward and scroll through history. Equally visible to you, equally revocable.
Invisible to the platform — these are grants macOS makes to OKEight. The system shows them to you; no chat platform is ever notified.
Provable with a packet capture — while reading your chats, not one packet goes to the chat platform's servers. Any network inspector confirms it.
Comparison
"Reading your screen" vs. "the way that gets you banned"
| Dimension | Reaching into the chat app | OKEight · reads the screen |
|---|---|---|
| Read method | Attach to process / decrypt database | System screenshot + on-device OCR |
| What it can reach | Your entire history (incl. what you aren't viewing) | Only what's shown on your screen |
| Platform detectability | High — process attach / fake client leave fingerprints | None — you look like a normal official-client user |
| Account risk | Documented bans | Triggers no anomalous platform-side signal |
| Where your data goes | Often uploaded / centralized | Stored locally; BYO key bypasses us entirely |