Pdfstudio

npx pdfstudio-mcp --root ./docs --read-only  # MCP: 10 PDF tools over stdio, no API keynpx pdfstudio info contract.pdf              # CLI — from a clone, after `npm run build:packages`

Quick start

In your browser, nothing to install: navigatorslab.com/pdf-studio

Self host with Docker, one command:

docker run -p 8080:80 ghcr.io/kayforkind/navigatorslab-pdf-studio:latest# → http://localhost:8080

CLI, from source (three commands):

git clone https://github.com/Kayforkind/NavigatorsLab-PDF-Studio && cd NavigatorsLab-PDF-Studionpm install && npm run build:packagesnpx pdfstudio info contract.pdf

The problem

Most "free" PDF editors work the same way: upload your contract or your ID to somebody's server, edit three pages, then get hit with a paywall, a watermark, a page cap, or a sign-up screen the moment you press Save. (The honest exceptions are self-hosted tools — but those make you run the infrastructure.) Desktop editors that genuinely modify existing text cost a subscription.

The fix

PDF Studio runs 100% in your browser. Your files are opened, edited, and saved on your own machine — there is no server to upload to. And unlike most free editors, it doesn't fake the editing: the Edit tool rewrites the page's content stream itself. The original text is deleted from the file and your replacement is written in as real vector text. Not an overlay. Not a text box on top. The old bytes are gone.

Try it in 30 seconds: open the live app, drop in any PDF, press E, click any line of existing text, retype it.

What it does

How it works

Everything happens in your browser — your file is opened, edited, and saved on your own machine. The pipeline:

1. Open — the PDF loads into browser memory. Nothing is uploaded.
2. Edit — the Edit tool rewrites the page's content stream: the original text is deleted from the file and your replacement is written in as real vector text.
3. Fill & sign — values go into the real AcroForm fields; signatures are drawn on the page. Flatten at export to lock them in.
4. OCR — scanned pages get a text layer from on-device Tesseract, making them searchable, editable, and redactable.
5. Redact — you draw boxes over what must disappear. In the editor these are marked for redaction (hatched): a promise, not a removal. Nothing is deleted yet.
6. Export — the file is rebuilt from scratch. This is where redaction destroys content: fully covered text is deleted from the content streams (recursively through nested form objects); anything that can't be proven removable (partially covered text, images, exotic fonts) forces the page to be rendered to pixels with the boxes burned in and the original content discarded; form values, annotations, accessibility-tree text, metadata, XMP, attachments, JavaScript, and thumbnails under the boxes are scrubbed. Then the finished file is scanned for every covered string across multiple encodings — if anything redacted is still recoverable, no file is delivered; export fails loudly instead of handing you a file that leaks.

After export the app shows a verification report (regions redacted, strings checked, recoverable: zero). Whiteout, by contrast, is visual-only — it covers content without removing it. Use redaction for sensitive data.

Full detail: How it works — docs

How it compares

PDF Studio is the zero-setup option that keeps a full editor and never uploads your files.

¹ Smallpdf, iLovePDF, PDF24 and similar. They differ from each other; the row reflects the category, not one product. ² Adobe's published plan, checked 2026-10-08 — prices and plans change.

Full, sourced breakdown: COMPARISONS.md · web version.

The developer difference: redaction here is not a promise, it's a gate. The exporter scans its own output for every covered string in UTF-8, UTF-16BE/LE and hex, through both the delivered bytes and a decompressed re-serialization. If anything is recoverable, no file is produced.

Honest caveats

What PDF Studio does not do yet:

• Text editing needs a real text layer. Scanned pages go through the on device OCR first.
• Replacement text uses Helvetica metrics, so on exotic embedded fonts the match is near identical, not glyph perfect.
• The on device AI runs small models (0.5B to 1B). It summarizes your document and answers questions about it well. It is not a frontier model.
• There is no real time collaboration and no cloud sync. Autosave lives in the browser you used.
• Signatures are drawn signatures placed on the page, not certified digital signatures.

Privacy — absolute, by architecture

Security

[Security: A-]

Security assessment — 2026-09-28 — overall grade A−. Each claim was checked against the source code. Full assessment: SECURITY_ASSESSMENT.md · agentic (CLI/MCP) addendum · reporting policy.

Residual risks (stated plainly): autosave is plaintext on your disk, so wipe after use on a shared machine; AI model weights come from HuggingFace over TLS without hash verification. Neither lets data leave your device.

Found something? Please report it privately — see SECURITY.md.

For developers

git clone https://github.com/Kayforkind/NavigatorsLab-PDF-Studiocd NavigatorsLab-PDF-Studionpm installnpm run dev        # http://localhost:5199npm test           # 88 app testsnpm run test:packages   # core, CLI, and MCP tests (builds packages first)npm run build      # → dist/ (relative base: works at domain root, subpath, or CDN)

Stack: TypeScript · React 19 · Vite · pdf.js (render) · pdf-lib (write) · Tesseract.js (OCR) · WebLLM (AI) · vite-plugin-pwa. Zero backend.

Three interfaces, one engine: the browser app, the pdfstudio CLI (9 commands), and pdfstudio-mcp (10 tools, publishable on its own — see packages/mcp). All three parse with pdf.js and rewrite content streams the same way; none of them make a network call with document bytes.

Bonus tool: packages/redact-check — drop a PDF and see the text still hiding under its black boxes. A small, self-contained proof of why "redaction" that only paints over text is worthless.

The interesting engineering: every mark lives in the page's PDF user space — the on-screen overlay and the exporter share the same CropBox-aware transform math (src/lib/viewport.ts), pinned in tests against pdf.js's own PageViewport, including offset MediaBox/CropBox origins that break most home-grown editors. Text edits go through a content-stream tokenizer: the original glyphs' show-text operators are deleted and replacements are written back with TJ kerning reproducing the original spacing. Tesseract, pdf.js workers, and WASM all run under a strict Content Security Policy with zero network calls for document data.

Test suite (88): i18n (language detection, switching, persistence, wipe, catalog completeness for es/de/fr/pt) · document model ops & undo semantics · viewport parity with pdf.js at 0/90/180/270° · export round-trips re-parsed with pdf.js · LCS diff engine · content-stream tokenizer + deep text-rewrite/vector-redaction round-trips · table-cell hit splitting.

Good first issues are labeled good first issue — see CONTRIBUTING.md.

Help translate

PDF Studio ships in English, Spanish, German, French, and Portuguese (Portuguese is a machine-drafted first pass awaiting native review), and the whole UI is localizable — the app bundles its translations, so switching languages works fully offline. If you speak another language, you can help:

• No code needed: translations are managed in Crowdin (link in the pinned issue) — translate or vote on strings in your browser.
• With code: edit src/locales/<lang>/translation.json (copy the English keys from src/locales/en/translation.json, keep every {{placeholder}} and <b>/<code> tag intact), then run npm test -- src/lib/i18n.test.ts — the catalog-completeness tests verify your file key-for-key against English.

New languages are picked up automatically once the JSON file exists and the language is registered in src/i18n.ts (LANGUAGES).

Agentic usage

PDF Studio ships two agent native interfaces that reuse the exact same content stream engine as the app. No new parsing code, no network calls, document bytes never leave the machine.

npm installnpm run build:packages

Run the CLI from the repo root with npx pdfstudio …, or make it global with npm link -w packages/cli and then call pdfstudio anywhere. The MCP server lives at packages/mcp/dist/index.js after the build.

CLI and MCP reference

Every command and tool is documented on the docs site: CLI reference · MCP reference.

Main CLI commands: pdfstudio info, extract-text, search, edit-text, redact, merge, split, rotate, pages. The MCP server exposes matching pdf_* tools. For least privilege, --read-only registers only the three read tools. Paths are realpath-resolved inside --root, and symlink escapes are rejected.

Conventions: - reads stdin, -o - writes to stdout, reports go to stderr. Exit codes: 0 ok, 1 error, 2 no matches or nothing changed. --json on read commands for scripting.

--read-only is the recommended default in the examples below. It registers only the three read tools (pdf_info, pdf_extract_text, pdf_search_text), so an agent cannot modify files. Drop the flag only when the agent needs the edit tools.

Claude Code:

claude mcp add pdfstudio -- node /path/to/NavigatorsLab-PDF-Studio/packages/mcp/dist/index.js --root /path/to/your/docs --read-only

Cursor / Cline / any MCP client (mcp.json):

{  "mcpServers": {    "pdfstudio": {      "command": "node",      "args": ["/path/to/NavigatorsLab-PDF-Studio/packages/mcp/dist/index.js", "--root", "/path/to/your/docs", "--read-only"]    }  }}

Privacy guarantee for agents: the CLI and MCP server make zero network calls. Same as the web app: parsing (pdf.js), editing (content stream rewriting), and writing (pdf-lib) all run in process. Redaction deletes the text operators from the file, and the output is rebuilt so orphaned bytes are gone too.

Prompt injection note: PDF content is untrusted input. Extraction and search results are delimited and labeled as data. Agents should treat document text as data, never as instructions. If your agent framework echoes tool output into its context, keep that boundary in mind.

Self-host with Docker

Zero backend, so self-hosting is one command. Images publish to GHCR on every release:

docker run -d -p 8080:80 --name pdf-studio ghcr.io/kayforkind/navigatorslab-pdf-studio:latest# → http://localhost:8080

Or build it yourself:

docker build -t pdf-studio .docker run -d -p 8080:80 pdf-studio

The container serves the static dist/ build behind nginx — your documents still never leave the machine running it.

Built by

Kazim Raza (@Kayforkind) — creator of PDF Studio and founder of NavigatorsLab.

FAQ

• Is it really free? What's the catch? — Free, MIT-licensed, no catch. No accounts, no watermarks, no caps, no "pro" tier. Fork it, self-host it, ship it.
• How is this different from other free PDF editors? — They let you annotate (text boxes over content). PDF Studio rewrites the content stream: original text operators are deleted from the file, replacements written in as real vector text.
• Honest limits? — Text editing needs a text layer (flat scans go through the OCR tool). Replacement text uses Helvetica metrics — near-identical, not glyph-perfect, on exotic embedded fonts. The on-device AI runs small models (0.5B–1B): great at summarizing your document, not a frontier model.

License

MIT — free for personal, commercial, and everything in between. That's the whole point: LICENSE.