pdfsmith 🔨

Every PDF tool, one tab. Merge, split, compress, organize, convert, protect — a fast, self-hostable alternative to the PDF-tools sites, with one big difference: it runs on your server, processes files in memory, and never stores anything.

pip install -e .
pdfsmith serve          # → http://127.0.0.1:8000

or, with every tool enabled (LibreOffice, OCR, Ghostscript included):

docker compose up --build

The tools

	Tool	What it does
🧩	Merge	combine PDFs in any order
✂️	Split	extract page ranges, or burst into one file per page
🗂️	Organize	reorder / duplicate / delete pages by dragging thumbnails
🔄	Rotate	all pages or just the ones you click
🗜️	Compress	lossless always; balanced/strong via Ghostscript
💧	Watermark	translucent text, diagonal or edges, tunable opacity
🔢	Page numbers	four positions, custom start, n / total format
✏️	Edit PDF	type text, stamp images, draw boxes — placed visually on the page
🖊️	Fill PDF form	fields appear as real inputs the moment you add the file
🔲	Crop	trim margins (mm) on all pages or a selection
🗑️	Remove pages	delete the pages you pick
🖼️	Images → PDF	JPG/PNG (alpha handled) into one document
📸	PDF → Images	crisp PNG/JPG at 72/150/300 dpi
📝	PDF → Word	editable .docx via pdf2docx
📽️	PDF → PowerPoint	one faithful full-bleed slide per page
📊	PDF → Excel	detected tables into .xlsx, one sheet per page
🗄️	PDF → PDF/A	ISO archival format via Ghostscript*
🌐	HTML → PDF	offline rendering, external resources disabled*
📄	Office → PDF	Word/Excel/PowerPoint/OpenDocument via LibreOffice*
🔍	OCR	make scans searchable via Tesseract + ocrmypdf*
📜	Extract text	the whole document as .txt
🔒	Protect	AES-256 password encryption
🔓	Unlock	remove a password you know
🖋️	Sign PDF	real PAdES digital signatures: your .p12, or a generated certificate
🕵️	Verify signatures	who signed it, and whether it changed since
⬛	Redact	black boxes that rasterize the page — the text underneath is gone
🕶️	Redact sensitive	auto-detect emails, phones, IDs, cards (+ custom terms) and black them out
⚖️	Compare	page-by-page text diff plus a visual difference score
🩹	Repair	rebuild damaged cross-reference tables
🏷️	Metadata	set title / author / subject
🧠	AI Summarizer	faithful Markdown summary of any text PDF*
🌍	Translate PDF	the document's text in another language*

…and ⚡ Workflows: chain tools into reusable pipelines ("Office → PDF, watermark, compress, merge") that run on any stack of files in one click — saved in your browser, free and unlimited.

* needs something from the operator: a system binary (LibreOffice, Tesseract, Ghostscript — all bundled in the Docker image), or for the AI tools an ANTHROPIC_API_KEY env var. The server detects what's present and the UI greys out missing tools with an honest hint; CI asserts every tool reports available inside the Docker image. The AI tools call the Claude API (claude-opus-4-8 by default; override with PDFSMITH_AI_MODEL) — they are the one case where document text leaves your server, and only when you enable them.

Why it feels fast

• No accounts, no upload-to-the-cloud round trip — your server is the cloud, and files live only in RAM for the duration of the request.
• A no-build frontend — vanilla JS + modern CSS, dark mode, drag-drop, keyboard shortcuts (/ to search), toast feedback. pdf.js renders page thumbnails client-side, so organizing pages is visual: drag to reorder, click to select, and the option fields fill themselves in.
• Honest errors — "page selection '99' is outside this document (1–2)" instead of a spinner that gives up.

API

The UI is a thin client over a clean API you can script against:

curl -F "files=@a.pdf" -F "files=@b.pdf" -o merged.pdf \
  localhost:8000/api/tools/merge

curl -F "files=@big.pdf" -F "level=lossless" -o small.pdf \
  localhost:8000/api/tools/compress

curl localhost:8000/api/tools        # tool list + options + availability

# chain tools server-side in one request — the Workflows engine
curl -F "files=@a.pdf" -F "files=@b.pdf" -o out.pdf \
  -F 'steps=[{"tool":"merge"},
             {"tool":"watermark","options":{"text":"FINAL"}},
             {"tool":"compress","options":{"level":"lossless"}}]' \
  localhost:8000/api/workflows/run

Errors come back as JSON with a real message; uploads are capped by PDFSMITH_MAX_MB (default 50). Workflow rules are enforced server-side: every step except the last must produce PDFs, only step 1 may take non-PDF input, and step errors report their position.

Architecture

pdfsmith/
  tools/
    registry.py      Tool dataclass: options schema, availability probe, runner
    basic.py         merge, split, organize, rotate             (pypdf)
    transform.py     watermark, page numbers, compress          (reportlab, pikepdf, gs)
    convert.py       images↔pdf, pdf→word, office→pdf, ocr      (pypdfium2, img2pdf, …)
    convert_more.py  html→pdf, pdf→pptx, pdf→xlsx, pdf→pdf/a    (weasyprint, python-pptx, pdfplumber)
    edit.py          visual edit overlay, crop, remove, redact  (reportlab, pypdfium2)
    sign.py          PAdES signatures + verification            (pyhanko, cryptography)
    forms.py         AcroForm listing + filling                 (pypdf)
    analyze.py       compare                                    (difflib, pypdfium2)
    secure.py        protect, unlock, repair, metadata          (pikepdf)
    redact_auto.py   auto-redact sensitive content (PII)        (pdfplumber, pypdfium2)
    ai.py            summarize, translate                       (Claude API, key-gated)
  workflows.py       the pipeline engine: validate + run step chains
  api.py             FastAPI: GET /api/tools, POST /api/tools/{name},
                     POST /api/workflows/run, POST /api/inspect/form-fields
static/              the frontend (no build step): tool grid + category tabs,
                     convert/all-tools menus, visual page editor, workflow builder
deploy/DEPLOY.md     Fly.io / Render / Railway / VPS+Caddy recipes

Every tool is a pure function (files, options) → result, registered with a schema the frontend renders automatically — adding a tool is one function and one register() call, and it appears in the UI with its options, no frontend work. Tools run on a worker thread pool, so a five-minute LibreOffice conversion never blocks the next user's merge.

Honest scoping

Where a capability has hard limits, the tool says so rather than pretending:

• Edit is an overlay editor (add text/images/boxes anywhere) — it does not rewrite existing content streams like a desktop editor.
• Redact rasterizes affected pages, which is why the hidden text is unrecoverable; redacted pages become images (run OCR after if you need them searchable).
• Sign produces real, validating PAdES signatures; a generated certificate is self-signed, so validators report it valid-but-untrusted unless you upload a CA-issued .p12.
• PDF → PowerPoint gives one faithful slide image per page; PDF → Excel extracts ruled tables it can detect and falls back to text rows.
• AI Summarizer / Translate send extracted text to the Claude API — they refuse scans (run OCR first) and oversized documents rather than silently truncating, and they're off until you provide a key.

Tests

107 tests cover every tool (AES and signature round-trips, tamper detection, redaction that actually removes text, hand-checked crop math, table extraction, compression that never returns a bigger file, workflow chaining and its validation rules, AI tools against a stubbed API — no network in tests) plus the full HTTP layer (error shapes, upload caps, the event-loop regression for signing). CI also boots the real server and merges PDFs over HTTP, then builds the Docker image and asserts all 32 tools report available inside it.

pip install -e ".[dev]"
pytest

Hosting

See deploy/DEPLOY.md — a $4 VPS with docker-compose and Caddy gives you HTTPS in ten minutes; Fly.io/Render/Railway recipes included.

License

MIT