Dominik Peters' TikZ Editor edits LaTeX figures visually without rewriting code

Editing a figure in a LaTeX paper has long meant hand-editing TikZ coordinates and recompiling to see the result. True WYSIWYG editing seemed structurally impossible: TeX is context-dependent and famously near impossible to parse, and the few visual tools that existed, such as TikZiT, could not import code that was already in a paper.

Dominik Peters' TikZ Editor sidesteps the 'parse all TeX' trap. It parses only the commands frequently used to make TikZ figures, builds a semantic scene graph that resolves coordinates, styles, transforms, loops and node anchors, and tags every element with the line and character ranges it came from. When you drag or resize a node, the editor changes the original source with a small patch rather than rewriting the file in a canonical format, so the user's indentation and line breaks survive, and it can explain errors without running a TeX compiler.

The project also shows what changed to make this feasible: competent LLM coding agents supplied the brute-force patience that parsing TeX had always demanded, and the initial code was written by Codex models over a three-month period. MathJax renders text and math, and the editor re-implements TeX's hyphenation and the Knuth-Plass line-breaking algorithm so multi-line labels look the way TeX would print them.

As of its June 2026 Hacker News launch the free, MIT-licensed editor sat at v0.5.2, running on the web and as a lightweight Tauri desktop app. It opens whole .tex papers with multi-figure previews, imports SVG, Ipe and PowerPoint files, exports to SVG, PNG, PDF or standalone LaTeX, and the launch discussion drew 451 points and 74 comments.

Parse the practical subset, then patch the original source in place

Scoping the parser to the commands figure-makers actually use turns a famously near-impossible problem into a tractable one.

Tagging scene elements with character ranges turns every visual edit into a surgical patch, so user formatting survives and 'bring your own code' becomes viable.

Keeping the TeX compiler out of the loop gives instant feedback and human-readable errors, which is what makes fine-tuning feel live.

LLM agents supplied the sustained brute-force work the parsing problem needed, collapsing the effort from impossible to a three-month build.

Free MIT licensing, web and desktop builds, and imports from SVG, Ipe and PowerPoint remove price and lock-in barriers to adoption.

For a format everyone calls unparseable, scope to what users actually write, keep your model linked to the original text, and edit with small patches rather than canonical rewrites.

By June 2026 the editor had reached v0.5.2: MIT-licensed and free on the web and as a desktop app built with Tauri, with support for opening full .tex papers, importing SVG, Ipe and PowerPoint files, exporting to SVG, PNG, PDF or standalone LaTeX, and an optional desktop assistant that edits figures through an installed Codex app server. The author published converter packages for svg2tikz, ipe2tikz and pptx2tikz as standalone npm libraries, and the Hacker News launch drew 451 points and 74 comments.