agent-pdf-cli

Which PDF framework should your AI agent use? We tested 8 so you don't have to.

---

Why This Exists

AI agents can write code. But when you ask one to generate a polished PDF -- with cover pages, tables, badges, dosing schedules, and pixel-matched typography -- most frameworks fight back hard.

We needed to know: which PDF framework produces the best results when the "developer" is an LLM?

So we took a real 4-page clinical protocol document and rebuilt it from scratch in 8 different frameworks, using AI agents (Codex, Claude) to write every generator script. Then we had Codex review each output against the HTML original.

This repo is the result. Every script, every output, every review -- all public.

Frameworks Compared

Framework	Language	Approach	Script
Typst	Typst	Native markup language	gen_typst.typ (v1-v5)
FPDF2	Python	Procedural API	gen_fpdf2.py
ReportLab	Python	Platypus flowables	gen_reportlab.py
WeasyPrint	Python	HTML/CSS to PDF	gen_weasyprint.py
Playwright	Python	Headless Chromium	gen_playwright.py
pdfmake	Node.js	Declarative JSON	gen_pdfmake.js
React-PDF	Node.js	React components	react-pdf/gen.mjs
LaTeX	LaTeX	TeX typesetting	gen_latex.tex

Results

Typst wins. After 5 iterations it produced the closest match to the HTML original. Its markup is clean, its output is sharp, and LLMs write good Typst code because the syntax is straightforward.

The short version:

• Typst -- Best overall. Clean syntax, excellent typography, fast compilation. LLM-friendly.
• ReportLab -- Strongest Python option. Verbose but precise pixel control.
• FPDF2 -- Lightweight Python. Good for simpler layouts, struggles with complex grids.
• LaTeX -- Beautiful output but the boilerplate is brutal. LLMs generate too many packages.
• pdfmake -- Decent declarative approach. Badge/circle rendering is limited.
• React-PDF -- Familiar if you know React. Limited layout primitives.
• WeasyPrint -- Great idea (just use CSS!) but CSS-to-PDF edge cases are painful.
• Playwright -- Screenshot-to-PDF. Least control, most dependency headaches.

Install

git clone https://github.com/199-biotechnologies/agent-pdf-cli.git
cd agent-pdf-cli

Python generators

pip install fpdf2 reportlab weasyprint playwright
playwright install chromium

Node.js generators

cd scripts && npm install

Typst

brew install typst  # macOS
# or download from https://github.com/typst/typst/releases

Quick Start

Each generator is standalone. Run any of them:

# Python
python scripts/gen_fpdf2.py
python scripts/gen_reportlab.py

# Node.js
node scripts/gen_pdfmake.js
node scripts/react-pdf/gen.mjs

# Typst
typst compile scripts/gen_typst_v5.typ output/typst-v5.pdf

All output lands in output/.

How It Works

data/protocol-data.json    <-- Sample protocol (patient data, peptides, supplements)
     |
     v
scripts/gen_*.{py,js,typ,tex}  <-- 8 framework generators
     |
     v
output/*.pdf               <-- Generated PDFs for comparison

1. Same input data -- every generator reads from protocol-data.json or embeds the same content
2. Same target design -- a 4-page Healtrix clinical protocol with cover, tables, badges, and monitoring grids
3. AI-written generators -- each script was written by Claude or Codex, not hand-coded
4. Codex reviews -- GPT-5.4 compared each PDF output against the HTML baseline

Project Structure

agent-pdf-cli/
  data/
    protocol-data.json       # v1 sample data
    protocol-data-v2.json    # v2 sample data (alternate schedule)
  scripts/
    gen_fpdf2.py              # FPDF2 generator
    gen_reportlab.py          # ReportLab generator
    gen_weasyprint.py         # WeasyPrint generator
    gen_playwright.py         # Playwright generator
    gen_pdfmake.js            # pdfmake generator
    gen_latex.tex             # LaTeX source
    gen_typst.typ             # Typst v1
    gen_typst_v2.typ          # Typst v2 (data-driven)
    gen_typst_v3.typ          # Typst v3 (Codex fixes)
    gen_typst_v4.typ          # Typst v4 (full rewrite)
    gen_typst_v5.typ          # Typst v5 (native design)
    react-pdf/gen.mjs         # React-PDF generator
  output/                     # Generated PDFs
  fonts/                      # Shared fonts

Contributing

See CONTRIBUTING.md. PRs welcome -- especially new framework generators.

License

MIT

---

Built by Boris Djordjevic at 199 Biotechnologies | Paperfoot AI