Best AI Agent Skills for PDF Processing in 2026

PDFs are where information goes to become hard to use. The text won't copy cleanly, the tables turn to mush, and the scanned ones aren't even text; they're pictures of text. Most "PDF tools" online solve exactly one of those problems and want your file uploaded to do it.

Agent skills take a different shape. Installed into Claude Code, they let an agent extract, convert, fill, and parse PDFs on your own machine, and chain those steps into real work. Here are the skills worth installing in 2026, what each does, and how to install them safely.

The short version

The core skill: pdf-tools

This is the one to install first. The pdf skill handles the full range of PDF work: pulling text and tables, merging and splitting files, filling forms, adding watermarks, encrypting, and running OCR on scanned pages.

What makes it more than a converter is chaining. Because Claude drives the skill, "extract the line items from these three invoice PDFs and total them by vendor" runs as one request. The agent reads each file, pulls the tables, and does the math, with no website round-trips and no copy-paste.

Install it from the official Anthropic repo:

git clone https://github.com/anthropics/skills.git
cp -r skills/document-skills/pdf ~/.claude/skills/

Restart Claude Code to load it. We've also got a step-by-step walkthrough in how to install the PDF skill in Claude Code if you want the longer version with troubleshooting.

On scanned documents and OCR

The part people underestimate: a lot of PDFs aren't text at all. A contract someone printed, signed, and scanned is an image. The pdf skill runs OCR, optical character recognition, to turn those images back into searchable, extractable text.

OCR quality tracks scan quality. A clean digital scan comes out near-perfect; a document that's been faxed, photocopied, and scanned at an angle will have errors. The rule that matters: spot-check OCR output before any of those numbers go into a report. The skill is fast, not infallible.

Conversions: docx-word and xlsx-spreadsheet

PDFs rarely stay PDFs. You extract from them to do something else.

The docx-word skill is what you pair with pdf when the destination is a Word document: pull content out of a PDF, rebuild it as an editable .docx with proper formatting. The xlsx skill is the partner for tables. Extract a table from a PDF and land it in a real spreadsheet with the columns intact, which is half the battle with financial documents.

Both are official Anthropic skills and install the same way as pdf above. Together with pdf they form a small document pipeline: read from one format, write to another.

In bulk: deep-research

When the job isn't one PDF but fifty (say, gathering and parsing a stack of filings or reports) deep-research helps on the front end. It runs multi-source web research and can collect the documents you then feed to the pdf skill for extraction.

It's an Anthropic skill, installed like the others. The combination turns "find and process every relevant report" into something an agent runs while you do other work.

Live reading: pdf-reader-mcp

For interactive work, pdf-reader-mcp connects Claude to PDFs through the Model Context Protocol, so the agent can read documents as part of a session rather than as a one-shot extraction. It's a community project, so the install is more involved and the security review matters more, so read the source before pointing it at sensitive files. If you're weighing this against the packaged pdf skill, our MCP servers vs agent skills guide covers the trade-off.

Installing safely

PDFs often carry sensitive content like contracts, financials, and personal data, so the security discipline here is stricter than usual.

The skills run locally and operate on files, which is the safe part. The thing to watch is what gets sent to the model for reasoning and whether a community skill makes network calls it doesn't need. Read the SKILL.md and scripts before installing, prefer the official Anthropic repository for sensitive work, confirm your Claude Code plan's data settings before processing confidential documents, and never let a PDF skill upload a file somewhere the task doesn't require. Our full agent skill security audit guide walks through the review step by step.

For the people who process PDFs all day (accountants, researchers, anyone buried in documents), see our full list for accountants for the GUI tools that complement these skills.

A real workflow: invoices into a spreadsheet

The most common request these skills handle is also the most tedious by hand: turning a folder of invoice PDFs into one clean spreadsheet.

Point Claude at the folder and ask it to use the pdf skill to extract the vendor, date, line items, and total from each file, then use the xlsx skill to write them all into a single sheet with one row per invoice. For digital PDFs this is near-instant. For scanned ones it runs OCR first, and that's where you slow down and verify: spot-check a few totals against the original images before trusting the sheet, because OCR can misread a smudged digit.

What used to be an afternoon of copy-paste becomes a few minutes plus a verification pass. The verification isn't optional, but it's far less work than the manual entry it replaces.

When not to reach for a skill

Skills aren't always the right call. For a single PDF where you just need to read it, opening it is faster than writing an instruction. For a one-time conversion of a non-sensitive document, a free online converter is fine. And for anything that needs pixel-perfect layout preservation, like a designed brochure, extraction tools (skill or otherwise) will fight you; you want the source file, not a parse.

The skills earn their keep on volume, repetition, sensitivity, and chaining. A single, simple, public PDF hits none of those. Match the tool to the job, and you'll know when to skip the setup.

On accuracy, and trusting the output

The honest limitation of PDF skills is that extraction is never 100% perfect, and the failure mode is quiet. A misread digit in an OCR'd invoice doesn't throw an error; it just sits in your spreadsheet looking like a real number. That's the trap, and it's worth naming plainly.

The defense is proportional verification. For digital PDFs with clean, selectable text, extraction is reliable enough that a quick scan of the result is fine. For scanned or low-quality documents, check a sample against the originals, more of them the more the numbers matter. For anything that feeds a financial decision, verify every figure, the same standard you'd apply to manual entry.

None of this makes the skills less useful. It makes them a tool with a known edge, and knowing the edge is how you use it safely. The time saved on extraction is real; spend a fraction of it on verification and you keep the speed without inheriting the risk.

FAQ

What can a PDF agent skill actually do? Extract text and tables, split and merge files, fill forms, add watermarks, and OCR scanned pages. Because Claude drives it, you can chain steps: "pull the line items from this invoice and put them in a spreadsheet" is one instruction.

Can skills read scanned PDFs? Yes, via OCR. Native PDFs are read directly; scanned ones are images converted to text first. Accuracy depends on scan quality, so spot-check the output.

Is it safe to process confidential PDFs? The skills run locally on files, which is safe. Confirm your plan's data settings, avoid free consumer tiers, and read the skill before running it on sensitive documents.

How is this different from an online PDF converter? A converter does one task and usually uploads your file to a server. A skill runs on your machine, handles many tasks, and chains into other work, which is safer and faster for anything sensitive or multi-step.