pdfops 0.4.0

One document on one machine, start-up included. Every row, the one pdfops loses too, is in the benchmarks.

Why pdfops

• Built for agents. Every command returns one JSON document, errors say what to do next, and text can be read under a character budget with a resume point.
• Everything in one place. Reading, layout, tables, OCR, rendering, page surgery, stamping, redaction, annotations, signatures, forms, encryption, inspection and creation: 31 tools behind one schema.
• Nothing to set up. A single binary with no PDF libraries, no Python and no runtime. Only OCR needs an extra program, and pdfops can fetch the language data itself.
• Fast. Files open in milliseconds whatever their size, and page work runs on all cores. See the benchmarks.
• Redaction you can trust. Content is deleted from the page, not covered, and the result is checked by a second interpreter before anything is written.
• Safe on files you did not write. Every command runs under a memory cap and a time limit, the MCP server runs each call in a process of its own and can be confined to one directory, and every command is exercised against a corpus of 988 hostile and malformed PDFs in CI.
• Knows what a file is up to. scan reports scripts, actions that fire by themselves, attached programs, disguised names and text that is extracted but cannot be seen, the carrier of prompt injection. sanitize removes the active content and proves it by scanning the result.

Quick start

As an MCP server, with nothing installed beforehand:

{
  "mcpServers": {
    "pdfops": { "command": "npx", "args": ["-y", "pdfops-cli", "mcp"] }
  }
}

Tools are named pdf_info, pdf_text, pdf_redact and so on, and take the same arguments as the CLI. Relative paths resolve against the server's working directory.

To keep an agent inside one folder, add --root: every path outside it is refused, including paths that arrive inside a document, and relative paths resolve against it.

{ "command": "npx", "args": ["-y", "pdfops-cli", "mcp", "--root", "/home/me/documents"] }

Install

The npm package is called pdfops-cli and installs the pdfops command. It is a small launcher: on first run it downloads the binary for your platform from the GitHub release, checks it against the published SHA-256 sum and caches it. Prebuilt binaries cover Linux (glibc and musl) and macOS on x86-64 and ARM64, and Windows on x86-64. The Docker image is Alpine with pdfops and tesseract with English; add -v for the folder to work in.

OCR additionally needs the tesseract program; nothing else does. Language data is fetched on request, without administrator rights:

pdfops ocr-langs                     # is tesseract there, which languages can be used
pdfops ocr-install --lang pol+eng    # download language data into the user's data directory
pdfops ocr-install --engine          # install tesseract itself where that needs no password

Tools

Run pdfops <command> --help for the options of each.

Examples

$ pdfops info manual.pdf
{"encrypted":false,"file":"manual.pdf","form_fields":0,
 "metadata":{"created":"2026-06-30T09:07:46+00:00","producer":"GPL Ghostscript 10.07.1"},
 "outline_entries":645,"page_size_pt":{"height":792.0,"width":595.0},"pages":357,"pdf_version":"1.3",
 "size_bytes":1386723,"uniform_page_size":true}

$ pdfops search manual.pdf "calling convention" --max-results 1 --context 40
{"file":"manual.pdf","matches":[{"match":"Calling Convention","page":12,
 "snippet":"... 10.5.1 The Pascal Calling Convention ..."}],"query":"calling convention",
 "total_matches":17,"unreadable_pages":[]}

Reading:

pdfops text manual.pdf --pages 12- --max-chars 4000     # resume_at_page says where to continue
pdfops text scan.pdf --ocr --ocr-lang pol+eng           # OCR only the pages that have no text
pdfops tables report.pdf --pages 4 --format markdown
pdfops layout report.pdf --pages 4 --level words        # bbox, font and size per word
pdfops render report.pdf --pages 1-3 --dpi 150 -o out/  # look at charts and layout
pdfops --stream ocr scan.pdf --lang pol                 # a line per page as it finishes
pdfops ocr scan.pdf --lang pol -o searchable.pdf        # the same file, with text to search
pdfops images report.pdf -o images/

Building and page work:

pdfops create notes.md -o notes.pdf
pdfops merge a.pdf b.pdf -o merged.pdf
pdfops pages in.pdf --keep "3,1,5-" -o out.pdf
pdfops split in.pdf --every 10 -o parts/

Editing:

pdfops stamp in.pdf --text "Confidential · {page}/{pages}" --position footer -o out.pdf
pdfops stamp in.pdf --image signature.png --x 380 --y 690 --width 140 --pages last -o out.pdf
pdfops stamp in.pdf --qr "https://example.com/doc/42" --anchor bottom-right -o out.pdf
pdfops redact in.pdf --text "Jan Kowalski" --text "\d{11}" --regex -o redacted.pdf
pdfops redact in.pdf --rect "2:100,200,300,220" -o redacted.pdf
pdfops replace in.pdf --find "2025" --with "2026" -o out.pdf
pdfops replace in.pdf --find "2025" --with "2026" --dry-run -o out.pdf   # the plan, nothing written
pdfops compress in.pdf --max-image-edge 1600 --image-quality 70 -o small.pdf

Forms and protection:

pdfops forms form.pdf
pdfops fill form.pdf --set name="Ada Lovelace" --set agree=true -o filled.pdf
pdfops encrypt in.pdf --owner-password secret --deny-copy -o locked.pdf
pdfops sign in.pdf --p12 identity.p12 --p12-password secret --reason "Approved" -o signed.pdf
pdfops sign in.pdf --cert me.crt --key me.key --visible "1:360,700,560,760" -o signed.pdf
pdfops signatures signed.pdf                            # valid, unchanged, who and when
pdfops signatures signed.pdf --trust company-root.pem   # and whether the signer is one of yours
pdfops sign in.pdf --p12 identity.p12 --tsa http://timestamp.digicert.com -o signed.pdf
pdfops scan inbox/offer.pdf                             # what is in it, before reading it
pdfops sanitize inbox/offer.pdf -o offer-clean.pdf      # scripts, actions, attachments removed

Markup:

pdfops annotate in.pdf --text "liability" --comment "check with legal" -o marked.pdf
pdfops annotate in.pdf --kind link --rect "1:72,50,300,70" --url https://example.com -o out.pdf
pdfops annotations marked.pdf

Conventions

• Output. Every command prints one JSON document on stdout. Errors go to stderr as {"error": "..."} with exit status 1. Add --pretty to indent.
• Progress. With --stream, ocr, text --ocr, render, images, split, redact and replace print one line of JSON per finished page or file, {"event":"progress","step":"render","done":7,"total":20,"page":12,...}, and the usual result as the last line. Pages are worked on in parallel, so events come in the order the work finishes; each names its page and done rises by one per line. Over MCP the same events arrive as notifications/progress when the call carries a progress token.
• Writing. Commands that write take -o. It may be the input file: output goes through a temporary file.
• Pages are 1-based and comma separated: 3, 2-5, 7- (to the end), -4 (from the start), 5-2 (descending), last, odd, even, all.
• Positions are in points with the origin at the top-left corner of the page as displayed, y growing downwards. layout reports them, stamp --x/--y and redact --rect accept them, and a pixel of render --dpi 72 is exactly one point.

How it compares

PyMuPDF is the closest in scope and is an excellent library; it is written in C, needs Python, and its AGPL license matters if you ship it. pdfops trades some breadth for a single MIT-licensed binary whose tools an agent can call directly.

Benchmarks

Best of 3 whole-process runs, start-up included, since that is what one tool call costs an agent. Document: the NASM 2.16 manual, 308 pages and 1.2 MB; images on a 352 page, 2.6 MB book with pictures; forms on a one page form. Machine: 8 core AMD Ryzen 7 9800X3D, Windows 11. Versions: poppler 25.07, qpdf 12.4, PyMuPDF 1.28, pypdf 6.19, pdfplumber 0.11, pyHanko 0.37 (CLI 0.5), tesseract 5.5. The fastest entry of each row is bold; n/a marks a tool that was installed and did not complete the task. PyMuPDF's entry for sanitize is its scrub.

Reproduce with scripts/bench.py on any document.

Why it is fast: info, layout, tables, render, images and ocr read objects on demand, so opening a file costs a few milliseconds whatever its size. Text extraction, layout, tables, rendering, splitting, image export and OCR run on all cores. Page operations copy only the objects the selected pages reach, so output size and time follow the selection, not the source.

Using it without MCP

pdfops tools prints [{"name", "description", "inputSchema"}] for every command. Pass these to any function calling API, then run the call through the CLI or the library.

let result = pdfops::tools::call(
    "pdf_text",
    serde_json::json!({"input": "report.pdf", "pages": "1-3"}),
)?;

Typed entry points live in pdfops::ops, for example pdfops::ops::read::text(TextArgs { .. }).

Behaviour worth knowing

pdfops <command> --help has the detail for each command.

• Any script. Text that stamp, fill, replace, create and ocr -o draw is shaped and embedded as font subsets; where no one font has every character, several share the text. Hebrew and Arabic are laid out from the right, Indic scripts and Thai are formed into their clusters, and all of it reads back as it was written, so a word is found by typing it.
• Replace writes in the document's own font where it has the glyphs and moves the rest of the line along; values of text fields and comments of annotations are changed too. A line that grows past its column passes its last words on to the next line, or is drawn narrower where it cannot. In a table all of this stays within the cell.
• Dry run. redact, replace, annotate and stamp take --dry-run: all the work, nothing written.
• Scan reports scripts, actions, attachments, disguised content and hidden text by severity, and with --clamav what ClamAV's signatures recognise, where that is installed. Sanitize removes the active content and scans the result before writing it.
• Signatures. sign appends, so earlier signatures stay valid. It can show the signature on a page (--visible) and embed a timestamp authority's statement of the time (--tsa). signatures checks that the bytes are unchanged and who signed; with --trust it checks the signer's chain against certificates you name, and with --revocation their revocation lists. Any other change to a signed PDF invalidates its signatures, as it must.
• Merging and page work keep bookmarks and links that lead to pages in the output, and rename the form fields of later inputs doc2.<name> so equal names do not share a value.
• Damaged files are repaired for reading and rebuilt for writing, encrypted ones too; the result then carries repaired_inputs. Protected files stay protected when edited; only decrypt removes the protection.

Limits

• Tables without ruling lines are inferred from how their text lines up, and say so (detected_by: alignment). An inference deserves a look.
• Replace draws a line no narrower than 70% of its width. What a replacement still runs over by is in overflow_pt, and --dry-run shows it before anything is written.
• Create follows the styling of HTML for how text looks: colour, background, bold, italic, underline, strike-through, size and alignment, from style attributes and <style> rules that go by tag, class and id. Layout is not read: no boxes, floats or grids.
• OCR is tesseract's, in quality and in languages.
• Scan is not a verdict: it never calls a file safe, and lists what it did not check.
• Resources. A command may use 4 GiB and 300 seconds by default: --max-memory, --timeout, or PDFOPS_MAX_MEMORY and PDFOPS_TIMEOUT; 0 lifts a limit. Rasters are capped at 64 megapixels.

Development

cargo test                                # under a second; fixtures are generated in memory
cargo test -- --ignored                   # OCR test, needs tesseract with a language pack
cargo clippy --all-targets -- -D warnings
cargo fmt --check
scripts/bench.py --help                   # regenerate the benchmark table
scripts/demo.py --help                    # record the session at the top again

Built on lopdf (object model), hayro (text, rendering, positions and image decoding), rustybuzz (text shaping), subsetter (font embedding) and pulldown-cmark (Markdown).

Contributing

Bug reports are most useful with the file that shows the problem, or the smallest one that still does. Pull requests need cargo fmt --check, cargo clippy --all-targets -- -D warnings and cargo test to pass, and a test for what they change: a fix comes with a test that fails without it.

License

MIT