A native Mac reader where you converse with what you're reading. The AI reads the document itself and cites the page. Built on Codex and Claude Code, so it runs on the ChatGPT or Claude plan you already pay for.
Johannes Vermeer, The Astronomer, 1668 · Musée du Louvre
what it is
Ask anything about the document, or select a passage and press Return to have it explained. The model looks up the pages your question needs and cites them; when a reply says “p. 12”, you click it and you're there.
the reading tool, not a chatbot with a file picker
The model has the document's table of contents and three tools: search its text, read its pages, and look at a page as it's printed, figures and tables included. It looks up what your question needs and cites the exact words; a click on a citation lands on them. Every step shows in the conversation. A short paper travels whole; a long book is read on demand.
The four tasks are Countdown, Math, SciKnowEval Chemistry and BoolQ (p. 3).
ES beats GRPO by the largest margin on Math: 74.0% vs 68.8%, a 5.2 point lead (p. 4).
Highlight a passage and it stays on the page in the color you chose, with a note if you like. Click it to ask about it. Export them all as Markdown.
Hold Option and drag over a figure or equation, then ask. And when a question is about a chart or a table, the model looks at the page itself.
Projects, documents, highlights, conversations, all local. The PDF's own contents in the sidebar, scanned pages read on your Mac, ⌘F that finds anything.
how it works
Internalize contains no AI of its own. It drives Codex, the agent inside OpenAI's ChatGPT app, or Anthropic's Claude Code, on your Mac and on your account. That's why it's free.
Internalize answers through the ChatGPT app for Mac or Claude Code, on the plan you already have. You sign in once, on OpenAI's or Anthropic's own page.
Open the DMG, drag to Applications, launch. Signed and notarized, no warnings, no setup, and it updates itself from then on.
Drop a PDF on the welcome screen, select anything, and ask. Your documents, annotations, and conversations persist in projects, locally.
Johannes Vermeer, The Astronomer (detail), 1668 · Musée du Louvre
privacy
What you ask, and the pages the AI reads to answer, leave your Mac only as your own requests through Codex or Claude Code. Delete a conversation and the copy they kept of it on your Mac goes too. Internalize itself only checks for updates.
We operate zero servers. Verify it with any network monitor.
You sign in on OpenAI's or Anthropic's own page, and Claude Code keeps its sign-in in your Mac's keychain. Internalize never reads it: it asks Codex or Claude Code who's signed in.
Each runs in Internalize's own isolated profile, with the shell, the browser and every other tool switched off, and none of your own settings loaded. They can read your document and search the web. They can't act on your Mac.
questions, answered honestly
Really free. The expensive part of AI apps is the AI. Internalize sidesteps that by running on the ChatGPT or Claude plan you already have, through Codex or Claude Code on your machine. There's no catch to monetize; if the app earns a place in your reading life, there's a tip jar.
Codex is OpenAI's local agent, and its app-server protocol is the documented way for third-party software to integrate it. That's what lets Internalize give you frontier-model conversations with zero API keys and zero per-token costs. Your plan, your machine, your data.
Yes. Internalize also answers through Claude Code, on a paid Claude plan. Open the Code tab in the Claude app once, or use Anthropic's installer, then sign in once: Claude Code opens Anthropic's sign-in page in your browser. It's a separate sign-in from the Claude app's, kept in Internalize's own private Claude Code profile. Use both if you like: the reply menu lists each one's models, any conversation can switch between them, and each conversation remembers its model. The document tools, page citations and web search work the same either way.
No. You sign in on OpenAI's own page (Internalize opens it for you) or in the ChatGPT app, and for Claude on Anthropic's own page, which Claude Code opens in your browser. Internalize never handles credentials; Claude Code keeps its sign-in in your Mac's keychain. It runs Codex in an isolated, read-only profile with no shell, and Claude Code in a private profile with only Internalize's document tools and web search, so your own settings are untouched and the app can't perform actions through either, only conversations.
Yes. VoiceOver announces each answer as it lands, reads its headings, and follows a citation to the words it quotes, and the app's own controls are labelled for it. From the keyboard, ⌘L puts you in the chat, Return asks, ⌘. stops an answer and Esc takes you back to the page. Reduce Motion and Increase Contrast are respected, and the chat has its own text size.
macOS 15 or newer, Apple Silicon and Intel (universal binary). The app is signed with an Apple Developer ID and notarized, and ships its own one-click updates.
The App Store sandbox forbids apps from launching external programs, which is precisely how Internalize drives Codex and Claude Code. Direct distribution with Apple's notarization gives you the same malware-scanned safety without that limitation.