To turn a ChatGPT or Claude answer into a proper Word document, copy it with the assistant's own copy button (which gives you Markdown), paste it into an editor that renders Markdown and understands the assistant's equation format, fix it there, then export to .docx. Pasting straight into Word skips the step that turns the formatting symbols into formatting.

Everyone who has tried it knows the result: hash signs in front of headings, asterisks around bold words, tables that are rows of vertical bars, and equations that arrive as \frac{a}{b}.

Yes — every AI feature in MindMax works with your own API key, with no account, no weekly allowance and nothing held back. The subscription exists to save you from managing keys and invoices, not to unlock the product. This post explains why, and what running on your own key costs you.

If you already pay for an API, being told to buy a second subscription before an app will talk to a model is a bad trade, and everybody involved knows it. You are being charged twice for the same tokens — once by the provider, once by the app for the privilege of sending them.

So MindMax does not do that. Add your own key and every AI feature works the same way it does for a subscriber: the same tiers, the same tools, no account, no weekly allowance. Not a reduced mode, not a beta path, not a setting that quietly disables the interesting half of the product.

This post is about why that is not generosity.

Asking AI about a screenshot of a chart works, but it throws away the text under the picture, the page it came from and the document around it. Boxing a region inside the document keeps all three, so numbers are read from the text rather than guessed from pixels, and the answer links back to the page.

Everybody has the same workaround. You hit a figure you cannot parse, you screenshot it, you paste it into a chat window, you type "what is this saying". It works often enough to become a habit.

It also fails in a specific, repeatable way, and the failure is not the model being bad at pictures. It is that by the time the picture arrives, three things that were sitting right next to it have been thrown away.

The simplest way to use AI with an Obsidian vault is to open the same folder in a second app that reads it in place. MindMax opens your vault directly — no import, no copy, no upload of the whole folder — and can answer questions across your notes and the PDFs and EPUBs sitting in your attachments folder, with quotes that open on the page. Obsidian keeps working on the same files.

Most AI options for Obsidian are plugins that send note text to a model. They help with notes. They do very little for the other half of a typical vault: the two hundred PDFs your notes are about.

MindMax v1.0 is the first release line after the public beta opened. The 1.0.1–1.0.7 builds that followed on the same day fixed the Windows installer; everything below is in 1.0.7. MindMax is still in public beta — the version number marks the update channel, not a finish line.

What is in it​

• Automatic updates. New versions download quietly in the background and install with one restart. Until now every update meant a fresh download from the website, which is how people end up three versions behind.
• Voice input. Dictate into the chat box and into quick notes instead of typing. Thinking out loud is often faster than writing, and the point of a quick note is that it costs nothing to take.
• Choose a model per task. Model settings are grouped into tiers, and each kind of job — answering, annotating, translating, writing — can be pointed at the tier you want. A fast model for quick questions and a stronger one for long work, instead of one model for everything.
• See what the AI is using. A usage view shows tokens and credits by task, so a long job that cost more than expected is visible rather than mysterious.
• A rebuilt Markdown editor. Tables you can edit like tables, with rows and columns you drag; lists you can fold; quotes and callouts that read cleanly; maths and Mermaid diagrams that preview as you type; document properties from the slash menu. The editor is where notes and drafts live, so it has to be pleasant before anything else on top of it matters.
• One search across everything. The search panel was rebuilt to search PDFs, books, Office files and notes in one place, with results grouped by passage rather than repeated per match.
• A book overview on the canvas. Open a book beside a canvas and get its structure laid out as a map you can expand from, instead of a summary you read once.
• Plans and sign-in in the app. Change your plan without leaving MindMax, and sign in with Google. Using your own API key still needs no account at all.

Known rough edges​

• Windows builds are not signed yet. SmartScreen will warn on first install — choose More info → Run anyway. macOS builds are signed and notarised, so they open without a warning. The installation guide walks through it.
• Two-column PDFs with figures in between remain the hardest extraction case. If a quote reads oddly, click through to the page.
• Voice with your own key needs a speech model. On the subscription it works out of the box; with your own key, transcription goes to a speech model at your provider — if it offers none, pick one for voice in model settings.

Where to start​

New here: open your first folder. Already using it: the update will offer itself — accept it and restart.

Every few days someone who has built their own retrieval setup asks the same question about MindMax, and it is a good question: if there is no embedding pass over my folder, how does it find anything?

The short answer is that it finds material the way you would — by structure, by the words on the page, and by trying other wordings — and that documents small enough to read are read rather than retrieved from. The longer answer is that this is a trade with a real cost, and pretending otherwise would be the kind of claim that falls apart the first time someone tests it.

MindMax's AI annotates a book sparsely on purpose: it marks only where a busy expert would actually stop, tells you up front how often it plans to stop, and two of its four lenses ask you questions instead of handing you conclusions. Fewer, better marks are the ones you reread.

The first reaction to an AI reading pass is often mild disappointment: it went through a whole chapter and left two notes. Surely it could find more than that?

It could. We decided it should not, and the reasoning is worth writing down, because "more annotations" is the kind of improvement that makes a product worse in a way users do not immediately attribute to the product.

To translate a whole EPUB with AI and keep it consistent, settle a glossary of recurring terms before translating anything, translate every chapter against that same glossary in one run, have the result checked for dropped passages and changed numbers, and read it side by side with the original. Pasting chapters into a chat window one at a time cannot do the first two.

Sentence quality stopped being the hard part of machine translation a while ago. Length is.

To OCR a scanned PDF into text you can rely on, recognise it once on purpose, check the result page by page rather than trusting one accuracy figure, re-run only the pages that came out badly, and keep the original scan untouched so every quote can be checked against the page image it came from.

A scanned PDF is a stack of photographs. Until the photographs become characters, nothing can search it, quote it or translate it — and an AI that answers questions about it anyway is answering from nothing.

To chat with multiple PDFs reliably, ask questions that name what you want compared, require every claim to say which document it came from and quote it, and look at which documents the answer actually read. The common failure is not a wrong sentence — it is a source that was never looked at, and an answer that sounds complete anyway.

One PDF is the easy case. The questions worth asking usually span ten: which of these papers use the method, where do these reports disagree, what does the contract say that the summary left out.