Papilot turns PDF annotations into AI edits to LaTeX source

On October 31, 2026, the author of an open-source tool called Papilot announced it on V2EX, pitching it as a way to edit LaTeX papers by annotating the compiled PDF rather than reading source code.

The workflow starts in the rendered PDF: you select a sentence or box off a formula and type what you want changed. Papilot locates the matching line in the source, bundles it with your comment, and hands the task to an AI model such as Codex, Claude, or Antigravity. When the edit comes back, each change is shown as a laid-out before-and-after comparison, with formulas rendered rather than shown as raw code. You can accept, reject, or comment on each one individually.

The author explained the motivation: he routinely uses AI to revise his LaTeX papers, but a git diff only reveals source-level changes and gives no direct mapping back to the PDF. Opening the rendered document afterwards, he could not tell which parts had been touched. What he wanted was a tool resembling Cursor, letting him review every change right inside the compiled document and judging the result from the PDF alone.

Until a review is complete, changes stay isolated in that task's own worktree. Only after everything has been reviewed can you choose to apply the edits to your working directory or commit them to a dedicated papilot/... branch, with the commit message written by AI. A built-in Git panel offers diffs, commits, and pushes as well. According to the announcement, Papilot runs locally or on a server. The only available report so far is this release post; there is no further material on subsequent developments.