<div align="center"> <img src="assets/banner.png" alt="beautiful-pdf-mcp — print-ready PDFs for AI agents" width="100%" /> </div>

📄 beautiful-pdf-mcp

MCP server that lets AI agents produce print-ready, typographically correct PDFs — magazine spreads, GOST lab reports, books, resumes — compiled by Typst, not an HTML export dressed up as a document.

🧠 The idea: the page is the unit of design

Most generated documents are built the lazy way: stack blocks on a canvas and let them fall where they may. The result is familiar — a heading stranded at the bottom of a page, an image floating in a half-empty sheet, a paragraph that trails off into nothing.

This engine thinks the way a magazine make-up editor does, one page at a time:

1. Budget first. Before any text is written, estimate_page_budget measures how many words fit one page of the chosen template — so content is written to size, not trimmed after the fact.
2. Fill the page like a block. Every page inside a continuous flow is composed to the bottom of the type area. The leftover text carries to the next page as plain prose — mid-sentence, even mid-hyphen, exactly like a printed book. Tails are normal; holes are not.
3. Verify, then ship. Every compile returns a per-page layout_report (fill %, holes, defects). The agent inspects rendered pages as PNGs and fixes problems before the user ever sees them. strict_layout: true refuses to produce a defective PDF at all.
4. Two-pass image placement. The compiler tracks the position of every paragraph, computes where each image actually lands, and recompiles with explicit placements — so a photo never tears the layout or strands itself on an empty page.

📸 Showcase

<div align="center"> <img src="docs/showcase_journal.png" alt="Journal template — editorial spread with wrapped images" width="760" /> <p><em>Journal: text wraps around photos, justified type with no rivers, the article flows page to page</em></p> </div> <div align="center"> <img src="docs/showcase_book.png" alt="Book template — A5, Van de Graaf margins" width="620" /> <p><em>Book: A5, mirrored Van de Graaf margins, chapter typography, illustrations interleaved into the story</em></p> </div>

Two of eight templates shown — run the test suite to render them all.

✨ Features

• 📐 Page-as-canvas engine — budget → compose → per-page QC → two-pass image placement. Every sheet is a finished block, not an accident.
• 🖼️ Magazine spreads — two photos placed diagonally with one continuous text threading around both, powered by meander; single photos get true side-wrap via wrap-it.
• 📏 Auto-fit single-page documents — a short resume or letter measures itself and scales typography up until the sheet is gracefully full.
• 🎓 GOST 7.32 compliance — sections on fresh pages, figures after first mention, full-width tables with captions above: Russian academic standards enforced structurally. Works in English and Russian (language: "en" switches Figure/Table/Contents labels).
• ✂️ No rivers, ever — justified text uses aggressive hyphenation costs so lines pack tight instead of stretching into word gaps.
• 🔁 Deterministic re-rendering — the document state lives in JSON; every edit re-lays-out the whole document by the rules, so nothing ever "drifts apart".

🚀 Quick Start

Prerequisites

• Python 3.10+
• Typst — brew install typst (or download a release)

Installation

git clone https://github.com/Kreminskaya/beautiful-pdf-mcp.git
cd beautiful-pdf-mcp
pip install -r requirements.txt

Connect to your agent

Add to your MCP client config (Claude Desktop: ~/Library/Application Support/Claude/claude_desktop_config.json, Cursor: ~/.cursor/mcp.json — same JSON for any stdio MCP client):

{
  "mcpServers": {
    "beautiful-pdf": {
      "command": "python3",
      "args": ["/absolute/path/to/beautiful-pdf-mcp/src/server.py"]
    }
  }
}

Restart the client — tools appear as beautiful-pdf__*. Then just ask your agent:

"Make me a magazine-style PDF article from these three photos and this text."

📚 Templates

| Template | Use case | Format | Body font | |---|---|---|---| | report | Business report, analytics | A4 | Source Serif 4 | | academic_ru | Thesis, lab report (GOST 7.32, en/ru) | A4 | PT Serif 14pt | | book | Long-form, fiction & non-fiction | A5 | PT Serif | | technical | API docs, developer guides | A4 | IBM Plex Sans | | portfolio | Portfolio, showcase | A4 | Noto Sans | | letter | Official correspondence | A4 | Source Sans 3 | | journal | Magazine / editorial layout | A4 | Lora + Cormorant | | resume | Modern two-column CV | A4 | IBM Plex Sans |

All 21 fonts ship with the repo — output is identical on every machine.

🛠️ How agents use it

budget = estimate_page_budget(template="journal", language="en")
# → words_per_page, lines_per_page: write the article TO BUDGET

doc = create_document(title="Between Rock and Sky", template="journal",
                      language="en", preset_overrides={"accent_color": "#c4a35a"})
sid = add_section(doc_id, "On Restraint", ARTICLE_WRITTEN_TO_BUDGET, level=1)["section_id"]
add_image(doc_id, sid, "photo1.png")      # photos embed into the running text
add_image(doc_id, sid, "photo2.png")      # second photo → diagonal spread

result = compile_preview(doc_id, pages="1-3")
# → PNG per page + layout_report: fill % and defects for every page

compile_pdf(doc_id, "article.pdf", strict_layout=True)
# refuses to ship a PDF with underfilled pages or layout holes

The loop is the point: budget → compose → look → fix → ship. The agent reads the layout_report numbers (words to add, lines short), inspects the rendered pages, and iterates until every page is a clean block.

<details> <summary>🧰 Full tool list (16 tools)</summary>

| Tool | Description | |---|---| | estimate_page_budget | Words/lines that fit one page of a template — call before writing | | create_document | Create a document, returns doc_id | | add_section | Add a section (Markdown content) | | update_section | Update a section's title or text | | remove_section | Remove a section | | add_image | Image with optional caption, width, position (auto, after:N, wraps, top) | | add_gallery | Grid of images | | add_table | Table with headers and rows | | add_code_block | Syntax-highlighted code | | add_callout | Callout box (info / warning / tip / danger / quote) | | compile_preview | Render pages to PNG + per-page layout_report QC | | compile_pdf | Final PDF; strict_layout=True fails on layout defects | | save_document / load_document | Persist / restore document state as JSON | | get_document_state / list_documents | Inspect session state |

</details> <details> <summary>🎛️ Per-document style overrides</summary>

Any preset key can be overridden per document via preset_overrides:

create_document(..., preset_overrides={
    "accent_color": "#2a9d8f",        # brand colour
    "page_num_position": "bottom-center",  # or top-left … bottom-right, none
    "header_rule": False,             # drop the thin running-header line
    "show_header_footer": False,      # no furniture at all
    "body_font": "PT Serif",
    "margin_left": "3.5cm",
})

</details> <details> <summary>📐 How the page engine works</summary>

Each template declares a page contract (data/styles.json): what a finished page looks like for that genre — fill thresholds, tolerated underfill, whether a final chapter page may end early (a book chapter can; a hole mid-article cannot).

Compilation is two-pass: pass 1 renders the document with invisible per-paragraph position marks, the server queries where every paragraph and image actually landed, computes optimal after:N anchors for position: "auto" images, and pass 2 recompiles with explicit placements. The QC (src/layout_qc.py) then grades every page against the contract and reports exact numbers — "page 3: 6 lines short, add ~40 words" — so the agent can fix layout arithmetically instead of guessing.

The full specification lives in docs/SPEC_PAGE_FILL.md (in Russian — the project's design constitution is CONCEPT.md).

</details>

🧪 Testing

python3 tests/render_all.py            # render every template to PNG + page QC
python3 tests/render_all.py journal    # one template
python3 tests/render_all.py showcase   # hand-finished showcase documents

Every page of every template is rendered to tests/output/ and graded by the page-fill QC — the test fails if any page violates its template's contract.

🗺️ Roadmap

• [x] 8 templates with shipped fonts
• [x] Page-as-canvas layout engine (meander spreads, auto-fit, GOST structure)
• [x] Page budget + per-page fill QC + strict_layout
• [x] Two-pass compilation with automatic image placement
• [ ] Bibliography tool with GOST citation style (gost-r-705-2008-numeric)
• [ ] Typst 0.15 upgrade (multiple bibliographies, variable fonts)
• [ ] Decorative drop caps for the book template
• [ ] More CV layouts (single-column, photo-left)

🧱 Tech Stack

| Layer | Technology | |---|---| | Typesetting | Typst 0.14+, meander, wrap-it | | Server | Python, FastMCP | | Layout QC | typst query position marks + per-page fill grading | | Imaging | Pillow (aspect detection, preview pipeline) | | Fonts | PT, IBM Plex, Source, Noto, Lora, Cormorant — bundled |

📄 License

MIT — see LICENSE.

---

⭐️ If this saves you from one more ugly AI-generated PDF — star the repo!