The pile of notes and the literature review matrix are the work between reading and writing, and the machine you write on can now do the first pass. Rectoly runs the AI on your own computer: it sorts your notes under headings without changing a word of them, and writes each paper up as a card whose every sentence points to a page. Nothing is uploaded — unless you choose to bring your own API key.
In Across papers, pick a collection or the whole library and click Organize with AI. The notes you wrote across the papers — comments on highlights, sticky notes, text boxes — are grouped under headings by topic. The AI writes only the headings; every note keeps its paper and page, and a click opens the paper there.
Save it as a research theme and open it beside the paper, like any other theme: rename the headings, drag a note somewhere better, add your own paragraphs, and export to Markdown or Word with a reference list — or, in the Zotero edition, to Zotero as a note. A note that fits no heading waits under a heading of its own.
Choose Paper cards… from the library’s menu, choose the papers, and they are carded one by one: research question, method, data, findings and the limitations the authors state, from each paper’s own text and in its own language, every sentence ending in its page. On Windows and Linux, the card of the paper you are reading sits in a tab of the side panel.
Open the table: one row per paper, five columns, a page on every sentence, and a click on a sentence opens the paper on the paragraph it cites. The same table is written to a folder as review-table.md and review-table.csv with one card per paper, ready for your spreadsheet or your notes. In the Zotero edition the pages are zotero:// links that open the PDF there, and the cards can be saved as Zotero notes.
The AI writes from the paper’s paragraphs; Rectoly then checks every sentence against them. A sentence with a claim or a number that is not in the paragraphs it cites is dropped, and the card says how many were. Stated limitations holds only what the authors themselves wrote. A field the paper does not state stays empty rather than guessed.
The same model as translation does the work, about 2.5 GB downloaded once: on Windows and Linux on your graphics card, on the Mac on Apple silicon. On a machine without the card for it, or if you want a stronger model, choose Your own API key in Settings › AI — OpenAI, Claude, Gemini, DeepSeek, OpenRouter, an OpenAI-compatible server, or Ollama or LM Studio on your own network. Rectoly says what will be sent and asks first.
Card the collection, open the table, and the first draft of your literature review matrix is on screen beside the manuscript — every cell leading back to a page.
The notes you left on the iPad or Android tablet are in Across papers on your computer too. Organize them here, and the theme is the outline you write from.
The model, the papers and the notes stay on this machine, so there is no third party to add to a data-management plan. If your lab has approved a provider, bring that key instead — and see exactly what is sent.
On Windows, Mac or Linux, everything is free for 14 days. Then it's a one-time purchase that works on every device signed in to the same account.