Make Codex read PDFs

Codex works in a terminal, and PDFs aren't text — so it can't read them. The fix is one shell command it can run itself: a zero-install CLI that converts any PDF (including scans, complex tables, and multi-column layouts) to clean markdown with a GPU vision-language OCR model.

$ npx pdftomarkdown document.pdf -o document.md

No install, no signup — works immediately anywhere Node 18+ and network access are available, including the Codex sandbox.

One-off: just tell Codex

Mention the command once in your prompt and Codex takes it from there:

> use npx pdftomarkdown to read report.pdf, then summarize the key findings

Permanent: add it to AGENTS.md

Paste this into your project's AGENTS.md (or the global ~/.codex/AGENTS.md) and Codex handles PDFs in every session:

## Reading PDFs

You cannot open PDF files directly. To read any PDF (local file or
https URL), convert it to markdown first:

    npx pdftomarkdown document.pdf -o document.md

then read the markdown file. GPU OCR takes 10-30 seconds per page and
cold starts can add a minute - always wait, never cancel the command.
For long documents, read the output file in sections. If the output is
truncated to page 1, the free demo tier is active; ask the user to set
PDFTOMARKDOWN_API_KEY for full documents.

Free API key for full documents

Without a key, the demo tier converts page 1 of any PDF — enough to try it. A free Developer key (100 pages/month, full multi-page, no watermark) takes 30 seconds:

Get a free API key with GitHub →

export PDFTOMARKDOWN_API_KEY=your_key_here

The CLI picks the key up automatically from the environment.