Features — one toolbox, three ways to use it

Every operation in the catalog exists as a web tool, an MCP tool for AI agents, and a REST endpoint — same behaviour, same free policy.

1 · Web tools: everything happens on your device

The website is a toolbox that runs in your browser using pdf-lib, PDF.js and Tesseract.js. Nothing goes outside: no upload, no account, no watermark, no “trial”. Start at the tool index — most people begin with PDF to Word, Merge PDF or Compress PDF.

2 · The MCP server: PDF tools for AI agents

Agents shouldn't screen-scrape PDF websites. The MCP server at mcp.pdfmcp.org/mcp exposes all 74 operations as typed tools with proper schemas, so Claude, ChatGPT, Cursor and friends can merge, split, OCR, convert, redact and extract directly. Setup takes about a minute per client — the MCP page has guides for 13 clients.

3 · The REST API: same tools, plain HTTP

POST api.pdfmcp.org/v1/jobs — submit, poll, download. The full operation catalog is machine-readable at GET /v1/operations. Walkthrough: the REST API quickstart.

The privacy model

Web tools never receive your file. Server jobs are inspected by a security gauntlet (antivirus scan, structural validation, active-content and resource-bomb checks) before processing, run in sandboxed containers with no internet access, and are deleted within an hour. Logs carry operation metadata only — never content, never file names. Full detail: Privacy Policy and Security.

The capacity model — free and generous

250 guaranteed operations per key per day, paced at 60 per hour; beyond that, a burst queue uses the machine's spare capacity for free. Hard caps (100 MB, 500 pages, 120 CPU-seconds per job) keep one user from starving everyone else. If the machine ever saturates, burst work politely waits instead of the service falling over. The whole policy, in plain language: Pricing.

Built for documents people actually have

Page surgery (organize hub), conversions (to PDF · from PDF), optimization, protection & signing, markup & forms, and extraction & OCR — including agent-first outputs like text as JSON with bounding boxes and tables as CSV, and privacy tools nobody else leads with, like stripping AI provenance metadata.