Copy from ChatGPT, paste into the converter, click DOCX. Headings, lists, code blocks, tables, and LaTeX equations all carry over. Same flow works for Claude, Gemini, and Perplexity.
pandoc -f gfm+tex_math_dollars+footnotes -t docx --mathml), with a fully-local docx-package fallback.$x$) and block ($$x$$) translate to native Office Math (OMML), fully editable in Word./api/convert-to-docx; temp files deleted after each request.ChatGPT, Claude, and Gemini all expose a small copy icon next to each response. Use it. Manual text selection silently strips markdown — code fences, list bullets, and LaTeX delimiters all vanish.
If you've already lost formatting to a manual selection, regenerate the response and copy with the button. There's no recovery path for stripped markdown.
Visit the homepage. No signup, no install — the converter loads as a single page and runs entirely in your browser.
Markdown Input is the default. Leave it on unless you're pasting raw HTML.
Cmd-V (Mac) or Ctrl-V (Windows/Linux). The right panel renders a paper-clean preview as you type — headings, lists, tables, syntax-highlighted code, and KaTeX equations all live.
The preview is what your DOCX/PDF will look like. If something's wrong here, fixing it in the markdown is faster than fixing it in Word.
Hit DOCX in the export bar. The file downloads instantly with a timestamped filename. Open in Word, Google Docs, Pages, or LibreOffice — all read it fluently.
Behind the scenes, the markdown POSTs to /api/convert-to-docx (server-side Pandoc) for highest fidelity — better tables, cleaner equations. Pandoc reads it from a /tmp file, produces the DOCX, the temp files get deleted, the binary streams back. If the server route is unreachable, the converter falls back to a fully-local DOCX path using the `docx` package. Both are solid.
Citations / footnotes
Perplexity and Claude sometimes emit numbered citations as Unicode superscripts or markdown footnotes. Footnote-style markdown (`[^1]`) round-trips cleanly; superscripts are kept as text but lose the link.
Image attachments
Inline images render in the export only if the source URL serves with permissive CORS. Generated DALL-E images via OpenAI's CDN do; private workspace uploads typically don't.
Tool-call traces
Function/tool blocks from agent-style chats render as code blocks. They look fine but lose the structural distinction from regular code.
When ChatGPT emits LaTeX ($E = mc^2$ or$$ \int_0^\infty e^{-x^2} dx $$), the DOCX exporter translates each equation to native Office Math (OMML). Word treats them as real equation objects — you can click in and edit any term, change the font, or copy them into another document. PDF and HTML render the same equations via KaTeX.
Paste a ChatGPT response and you’ll have a clean DOCX in under five seconds.
Open the converter