← SistineGPT

Documentation

Everything in the app and how to use it. Machine-readable version at /llms.txt.

SistineGPT is a private chat for the Mac: a ChatGPT-style chat in one ordinary window, with a vault that is encrypted on your Mac, a choice of a model on your Mac (free) or private cloud AI, sealed end to end (Private Cloud), and a Ghost mode that saves nothing. Requires macOS 14 or newer on an Apple silicon Mac.

Contents

• Getting started
• The window
• Local and Private cloud
• Chats
• Ghost mode
• Voice
• Bring your own AI
• Abilities and approvals
• Appearance and instructions
• Privacy
• Keyboard shortcuts
• FAQ

Getting started

Download the app, drag it to Applications, and open it. A short setup walks you through the essentials: how to bring SistineGPT up from any app, voice permissions (skippable), and a one-click Set Up Free Local AI button that downloads a model sized to your Mac's memory from Ollama's library and runs it on your Mac. When it finishes, chat works with no account and no key.

Pricing is shown plainly during setup, before any paywall: Free is the product, not a teaser. Private Cloud ($19/mo) adds private cloud AI, sealed end to end. To buy it, open Settings → Plan and choose Select beside Private Cloud; Paddle, our merchant of record, takes the payment in your browser.

The window

SistineGPT lives in one ordinary Mac window. It is in the Dock, in ⌘ Tab and Mission Control, and it goes full screen like any app.

• The sidebar on the left: the ghost beside "SistineGPT", then the mode and the model ("Local · Qwen 2.5", "Network · GPT-OSS" or "Private cloud · GLM 5.3 Flash") as the menu that picks Local or Private cloud, then New Chat, a search field, your recent chats newest first, and Settings at the bottom. A model on your network wears an "Unencrypted on your network" badge under the model; a service you added wears "Not private". Past the month's allowance the label reads "Limit hit · Local Qwen 2.5" while replies come from the model on your Mac. The sidebar is open by default; ⇧⌘S or the button in the title bar hides and shows it.
• The title bar holds the traffic lights and the sidebar button, nothing else; the empty strip moves the window.
• An empty chat shows the ghost, and the composer sits at the bottom of the pane: the Ghost mode button and attach on the left, the microphone and Send on the right.
• Open and close it like a messaging app. Closing the window (the red button or ⌘W) keeps SistineGPT running, so a reply that is still streaming finishes and notifications still reach you. Click the Dock icon to bring the window back; ⌘M minimizes it; ⌘Q quits.
• Resize and move it like any window. It reopens at the size and place you left it.
• Drop a file on the chat (text, Markdown, RTF, PDF, or an image) to attach it.
• Global shortcut: press ⌥ Space (configurable) from any app to bring SistineGPT to the front, and again to hide it.
• Menu bar icon: Open SistineGPT, push-to-talk, Summarize My Screen, Ask About Clipboard, New Chat, New Ghost Chat, Lock (while the lock is on), Check for Updates, Settings, Launch at Login and Quit. It is on by default; hide it in Settings → General.
• Notifications: a reply that finishes while the window is not in front shows a notification; clicking it opens the chat.

The ghost renderer derives from “Ghost” by OpenProcessing user “me” (openprocessing.org/sketch/1138728, CC BY-SA 3.0), substantially reworked; the artwork is licensed CC BY-SA 4.0.

Local and Private cloud

One choice for the whole app, made on the mode menu: where the next reply runs. Every label reads the same switch, so they cannot disagree, and every reply's footer says where it actually ran: "Local · Qwen 2.5", "Network · GPT-OSS" for a computer on your network, "Private cloud · hardware verified" or "Not private · host".

• Local is a model on your Mac or on another computer on your network, or a service you added yourself. It is the whole of Free. A chat with the model on your Mac never reaches our servers. The assistant's tools live here: web search and web pages, each behind an approval card, memory notes, and your screen when you ask.
• Private cloud is the mode included in the Private Cloud plan. It runs on a sealed server that Venice routes to. Before sending, your Mac checks that the server is genuine sealed hardware and encrypts your message to it. It does not check every program the server's operator runs. The reply streams back encrypted and is decrypted on your Mac. Our relay carries ciphertext it has no key to open. A Private cloud turn carries text only: your message, recent messages from the chat and the text of a document you attached. It never carries your memory notes, your screen text or an image, and no tool is offered to it.
• Private cloud fails closed. If the server's hardware proof does not check out, Private cloud is switched off, the menu says why, and that reply comes from the model on your Mac with one line saying so, or is refused when no model is set up on your Mac. Nothing goes to the server in the clear instead. Later replies run where Local runs for you, on your Mac or on the computer on your network you chose, and the menu says which. Picking Private cloud again is the only thing that tries the server again.
• If the relay is not sealed. The app takes the sealed path only while our relay reports that it carries sealed conversations. If it ever stops, Private cloud mode cannot be picked, a Private Cloud subscriber's chat runs on their Mac or their network, and Settings → AI and the Plan tab say so in one line.
• At the cap, when the month's allowance is spent, chat falls back to the model on your Mac rather than stopping; the mode label reads "Limit hit · Local Qwen 2.5" until next month.
• View hardware check under a Private cloud reply shows what your Mac checked, while the Proof panel still holds the proof. Settings → Privacy → Proof exports it so anyone can re-check it.

Chats

Ask anything in the composer. Replies stream in with full Markdown and code blocks you can copy.

• Attachments: the + menu attaches documents and images; you can also drag files in or paste an image with ⌘V. A document is read on your Mac and its text goes with your question; the file itself is never uploaded. An image is pixels, and goes only to a vision-capable model on your Mac or your network, in either mode. What that model writes about the picture stays in the chat, so the rest of that chat also runs only on your Mac or your network.
• Summarize My Screen takes one snapshot when you ask, converts it to text on your Mac, throws the image away, and attaches the text. That text goes only to a model on your Mac or your network. Ask About Clipboard (⇧⌘V) attaches what you copied.
• Chat history is the sidebar, searchable, with retention you control in Settings → Privacy (keep forever, or 90, 30 or 7 days).
• Memory: the assistant can keep short notes between conversations, one list. Each note it saves appears in the chat with Undo. Notes ride along only when the model runs on your Mac or your network, never on a Private cloud turn and never in Ghost mode, and the assistant never saves one in a turn that read a web page, a document, an image or your screen. The notes are encrypted on your Mac; Settings → Abilities → Memory is where you read, edit and delete them, and Settings → Privacy clears the list at once.
• Instructions: one custom system prompt, under Settings → AI, stored encrypted in the vault with a Reset to the default. It goes with every request, in either mode, so anything you put there reaches the model.

Ghost mode

The ghost button on the composer is Ghost mode, like a browser's incognito window: a setting on the chat, not a kind of chat. Turn it on for a new chat (or press ⇧⌘N, New Ghost Chat, in the File menu or the menu bar icon) and the chat is written to no file on your Mac: it lives in a store that exists only in memory for the life of the chat. It reads no memory notes and saves none. Turn the button off on a chat that has messages and SistineGPT asks first; choose Save This Chat and it becomes an ordinary saved chat from then on, in place, while Stay in Ghost Mode writes nothing. A chat that was already saved cannot be made unsaved, so the button is disabled there. Rest the pointer on the button and a small card says what it does. Leaving a Ghost chat for another chat, closing the window or quitting asks once, if it has messages: Drop, Keep, or Cancel. It runs Local or Private cloud like any other chat.

Voice

One microphone button, two idioms: hold it to talk walkie-talkie style (release to finish) or click to latch recording until you click again. Esc cancels. A status row shows "Getting ready…" while the engine warms up, then "Listening, go ahead" with a live transcript as you speak, so no words get lost. Push-to-talk also has a global shortcut (⌥ V by default).

Apple speech recognition transcribes your voice on your Mac. If your language would need Apple's servers, SistineGPT tells you so and does not send the recording. The app records only while you hold the button (or, if you click to latch, until you click again or press Esc), and never listens in the background. SistineGPT can also speak responses aloud, a setting under Settings → Voice, with a choice of system voice, speaking rate, and an option to skip code blocks.

Bring your own AI

Use our inference or yours. SistineGPT is the client and the vault. Four ways in:

Source	What it is	Cost
Local	One click sets up a model on this Mac, sized to its memory, bound to localhost. Your conversations with it stay on your Mac.	Free forever
Another computer	Point the app at a Mac on your network running Ollama or LM Studio. Discovered automatically, picked from what is actually installed there. Requests travel over your network without encryption, and the app says so beside the model at the top of the sidebar.	Free
Private cloud	Private Cloud includes a model on a sealed server with no setup: each message is encrypted on your Mac to that server, after your Mac checks that its hardware is genuine.	Private Cloud
Your own service	On the Private Cloud plan, Local mode can point the app at any OpenAI-compatible endpoint of your own, with your own key, kept in the macOS Keychain. That service receives what you send, the app says "Not private" beside the model, and Sistine's privacy guarantees do not cover it.	Your own rates

The app is smart about local models: it detects Ollama and raises the context window so long prompts don't get silently truncated, and routes an image question to a vision-capable model when your main model cannot see. On capable Macs, Settings → AI offers an image model to download, with its size; nothing downloads until you press the button. For local servers, a capability line under the model picker tells you what your choice can do before anything fails mid-conversation.

Free is local and network only by policy. Private Cloud is capped by dollars, not message counts; at the cap, chat falls back to your local model rather than stopping.

Abilities and approvals

Abilities are what the assistant can do beyond talking, grouped under switches in Settings (memory, web access, screen awareness, desktop actions). Memory and web access start on; screen awareness and desktop actions start off. Tools you haven't enabled are never even shown to the model, so it can't be talked into trying them, and no tool at all is offered on a Private cloud turn.

• Shown as it happens, with Undo: a note the assistant remembers. Forgetting several notes at once asks first.
• Always asks first: web searches (the card shows the exact words, and Allow for this chat covers the rest of that chat until you quit), opening links, adding calendar events, running your Shortcuts, and reading your screen. The approval card names the exact action and target, "Not now" sits first, and pressing Enter does nothing, so you can't approve by reflex.
• Fails closed: background runs cannot approve anything, so searches and side-effecting tools simply don't run unattended.
• Untrusted content is fenced: anything read from the web, a document, or your screen is wrapped in markers the model is told not to obey. A page address you didn't type and no search returned needs approval before it's read, and a page from search results asks too when the chat holds your notes, a document, or your screen. While such content is in a turn, the assistant cannot save notes.

Web search is included with zero setup in Local mode: 40 searches a day, on Free and on the paid plan alike, through Sistine's search service to Brave Search, no account. On Private Cloud those searches also count toward the month's allowance. Add your own free Brave key to make searches unlimited, take them off the allowance, and send them to Brave directly, without our server.

Appearance and instructions

Four looks: Match System, Light, Dark, and Gallery, which hangs a public-domain masterpiece behind the window's dark glass. Ten paintings ship with the app (Van Gogh, Vermeer, Hokusai, Monet, Turner, and friends), rotating on a half-hour clock. Pin the one you love, or add your own images and they join the rotation.

You can also change the chat text size and reduce animation, and write the assistant's Instructions under Settings → AI.

Privacy

The short version. The privacy policy lists every service the app contacts, and when. What we prove, and what we promise sorts these claims into the ones you can check yourself, with the commands, and the ones that rest on our word. Settings → Privacy says the same in the app: where a document and a message go, what Sistine, Venice and the model can each see, and the metadata no setting hides.

• With a model on your Mac, your conversations stay on your Mac. A web search leaves only after you approve its exact words. Update checks, model downloads, and the features you turn on use the internet as the policy describes.
• In Private cloud mode, on the Private Cloud plan, your message, recent context from that chat and the text of a document you attached go to a sealed server that Venice routes to. Before sending, your Mac checks that the server is genuine sealed hardware and encrypts your message to it. It does not check every program the server's operator runs. Our relay carries ciphertext it has no key to open; the sealed server holds the plaintext of that turn while it writes the reply. The proof page says what that does and does not cover.
• Your screen text, memory notes and attached images go only to a model on your Mac or your network, on every plan and whatever the settings say. A chat bound past your network is not offered the tools that read them and carries no memory notes; work that needs them waits for a local model instead of going out.
• Conversations, memory notes, document records and your Instructions are saved on your Mac and encrypted there, with a key held in your Keychain that never leaves the Mac. Settings → Privacy can put that key behind Touch ID, which closes your conversations and notes while the Mac is locked. A chat in Ghost mode is written to no file unless you keep it.
• Screen reading takes one snapshot when you ask, converts it to text on your Mac, and throws the image away. Voice is transcribed on your Mac, and the microphone is on only while you use voice input. Crash reports are not sent anywhere.
• API keys live in the macOS Keychain. Updates are signed and verified before they can install. Settings → Privacy deletes everything SistineGPT saved on this Mac, your memory notes included, and destroys the key it was encrypted with.

Keyboard shortcuts

Shortcut	Action
⌥ Space	Bring SistineGPT to the front from any app, or hide it (configurable)
⌥ V	Push-to-talk from any app (configurable)
⌘N	New chat
⇧⌘N	New Ghost chat
⇧⌘S	Show or hide the sidebar
⇧⌘V	Ask about clipboard
⇧⌘L	Lock (while the Touch ID lock is on)
⌘W	Close the window (SistineGPT keeps running)
⌘M	Minimize to the Dock
⌘0	Bring the SistineGPT window back
⌘+ / ⌘-	Bigger or smaller chat text
Esc	Cancel voice
⌘,	Settings
⌘Q	Quit

FAQ

Is SistineGPT free?

Yes. The app runs free with a one-click model on your Mac, forever, including 40 web searches a day. Private Cloud ($19/mo) adds private cloud AI, sealed end to end, with a monthly allowance that falls back to the model on your Mac at the cap. Those are the two plans we sell; the earlier Search, Pro and Max plans are closed to new subscriptions, and the people already on them keep them.

Does SistineGPT send my data anywhere?

With a model on your Mac, your conversations stay on your Mac, and a web search leaves only after you approve its exact words. In Private cloud mode, on the Private Cloud plan, each message is encrypted on your Mac to a sealed server that Venice routes to, after your Mac checks that its hardware is genuine, and our relay carries ciphertext it cannot open. Your Mac does not check every program the server's operator runs. Your screen text, memory notes and attached images go only to a model on your Mac or your network either way. The privacy policy lists every service the app contacts.

What is Ghost mode?

The ghost button on the composer. On, the chat is written to no file on your Mac and reads no memory notes. Leave it and it is gone, unless you keep it, which copies it into your history.

What Macs does it run on?

macOS 14 Sonoma or newer, on an Apple silicon Mac (M1 or later). Intel Macs are not supported: the app ships as an arm64 build and will not launch on them.

What can the assistant actually do on my Mac?

Only what you enable, and web searches and anything with a side effect (opening links, calendar events, Shortcuts) show an approval card naming the exact action first. Background runs fail closed.

Questions this page doesn't answer: support@sistine.ai