read-paper

Citing a PDF paper inside a DeepSeek Harness conversation has usually meant screenshotting figures, copy-pasting fragments, and bouncing between a browser and a terminal. read-paper removes that friction at the source: running on the host via poppler, it rebuilds a two-column PDF into clean, structured text with headings, tables, sub/superscripts, and a dot-leader table of contents each in their proper place; figures and display equations are rendered as cropped bitmaps while the equation source stays searchable in hidden spans.

In the reading panel the paper occupies a dedicated third column beside the chat, collapsing gracefully into a floating panel or a slim tab on narrow windows. Selecting any passage summons a small ask dialog; type a follow-up question, and the selection, its surrounding context, and the abstract are sent to the model together. One click on Create project materializes the paper (PDF, Markdown, figures, metadata) into a Paper workspace and opens a session with the paper already loaded as context, so multiple papers can coexist in parallel.

It is aimed at researchers, graduate students, and engineering teams who regularly discuss papers with an AI assistant: instead of shattering a PDF into snippets and pasting them one by one, you read the original text and ask targeted questions on a specific derivation or figure without ever leaving the conversation.