smart-ocr Skill (claude-office-skills) for Claude Code: What It Does and How to Install It

Makes your agent pull text out of images, screenshots and scanned PDFs by writing and running PaddleOCR code.

Installs: 8.7K; on skills.sh
Stars: 556; on GitHub
Updated: Jan 2026; last repo commit

smart-ocr is a knowledge skill for PaddleOCR. It shows the agent how to load an image or a scanned PDF, pick a language, run OCR and return text with box positions and confidence scores. SKILL.md includes patterns for preprocessing images, rebuilding line layout, batch jobs and three worked examples: a business card reader, a receipt scanner and a multi-language document.

The agent has to run Python with PaddleOCR installed, which SKILL.md says to set up with pip. The frontmatter description is only the character >, so name the skill in your prompt if your agent does not pick it up.

Good for

People digitizing scans, screenshots, receipts or business cards
Developers who need text with bounding boxes and confidence scores
Anyone reading Chinese, Japanese, Korean or other non-English documents

Install smart-ocr

Claude Code

Copy this prompt.

Install the smart-ocr skill in Claude Code. Run this command:

npx skills add https://github.com/claude-office-skills/skills --skill smart-ocr

Then check that it loaded and tell me how to use it.

2
Paste it into Claude Code. It shows the change and asks before running it.

Requires Node.js.

The skill runs Python with PaddleOCR packages that you install with pip, so there is no Claude app upload route. See Needs in the facts box for the packages.

Built from the package data we hold. Not yet tested by us. Source.

What's in the skill

When it loads

>

Sections in SKILL.md

Overview
How to Use
Domain Knowledge
PaddleOCR Fundamentals
Supported Languages
Configuration Options
Processing Different Sources
Result Processing
Text Layout Reconstruction
Best Practices
Common Patterns
Image Preprocessing
Batch OCR with Progress
Examples
Example 1: Business Card Reader
Example 2: Receipt Scanner

Key facts

Author: claude-office-skills (named in the SKILL.md frontmatter)
License: MIT (SKILL.md frontmatter and the repo LICENSE)
Installs on skills.sh: 8,666 (Nov 2, 2026)
GitHub stars: 556 for the whole claude-office-skills/skills repo (Nov 2, 2026)
Last repo commit: Jan 31, 2026
What it includes: 1 file: SKILL.md
Audits on skills.sh: Gen Agent Trust Hub: pass, Socket: pass, Snyk: pass (Nov 2, 2026)
Needs: Python with paddlepaddle, paddleocr, pdf2image and Pillow, per the Installation section
Part of: claude-office-skills/skills: the README says 136+ skills, and the repo tree has 137 skill folders

What it includes

Loads PaddleOCR with angle classification and returns box, text and confidence
The result structure is a list of box, (text, confidence) pairs.
Runs code · Basic OCR
Uses a lang code, with 13 codes listed and 100+ languages claimed
Examples: en, ch, cht, japan, korean, french, german, spanish, russian, arabic, hindi, vi, th; multilingual is also shown.
Runs code · Language selection
Documents detection, recognition, angle, GPU and CPU settings
Includes det_limit_side_len, det_db_thresh, use_gpu and enable_mkldnn.
Configuration options
Runs OCR on one image, a list of images, or bytes
Single and multiple image examples.
Read · Image files and bytes
Converts pages with pdf2image, writes a temp PNG per page, OCRs it and deletes it
Joins the pages with page separators.
Runs code · Scanned PDFs
Downloads an image with requests and OCRs it from memory
SKILL.md's example fetches a URL.
Network · URLs
Sorts boxes by position and groups them into text lines
Uses a line threshold of 10 in the example.
Layout reconstruction
Grayscale, contrast and sharpen first; thread pool with a progress bar
The batch example uses 4 workers by default.
Runs code · Preprocessing and batches
Business card reader, receipt scanner, multi-language document
Regex-based parsing of email, phone and totals.
Worked examples

From the official docs

Permissions

Medium risk

What it can do

Read your data
Create and change things
Delete things
Read-only mode

Watch out for

It installs Python packages with pip
The Installation section runs pip install for paddlepaddle (or paddlepaddle-gpu), paddleocr, pdf2image and Pillow. Use a virtual environment so the packages stay separate from your system Python. github.com
The URL example downloads from the network
One pattern fetches an image with requests.get before OCR. Only point it at addresses you trust. github.com
The frontmatter names an MCP server
The frontmatter lists server office-mcp and tool extract_text_from_pdf. The rest of SKILL.md never calls them. The repo README describes office-mcp as a separate server you build with npm, so you do not need it for this skill's examples. github.com

Technical details

Turning on read-only	No read-only mode. The PDF example writes and deletes temporary page images.
Scopes	Runs Python code that reads the image and PDF files you point it at Installs Python packages with pip Writes temporary page images while OCRing a PDF

Questions

What does smart-ocr do?

It makes your agent extract text from images, screenshots and scanned PDFs using PaddleOCR. It can return the text with box positions and confidence scores.

Why does its description show only ">"?

The frontmatter description of SKILL.md is the single character >, so it gives an agent no text to match a request against. Our advice: name the skill in your prompt (for example "use smart-ocr to read this scan") if it does not load on its own.

How do I install it?

Use the install box at the top of this page for the skill. Then install PaddleOCR itself, as the Installation section of SKILL.md describes. There is no Claude app route, because the skill depends on local Python packages.

Does it need an API key?

SKILL.md names no key or account. It does need Python packages installed with pip, and it assumes your agent can run Python.

Which languages does it read?

SKILL.md says PaddleOCR supports 100+ languages and lists 13 codes, including English, Chinese (Simplified and Traditional), Japanese, Korean and Arabic. It also shows a multilingual setting.

Is it accurate on handwriting?

SKILL.md's own Limitations list says handwritten text accuracy varies. It also lists very small text, complex backgrounds and rotated text without angle classification as weak points.

Is it safe?

It runs Python that reads the files you give it and installs packages with pip. Our advice: use a virtual environment, and check receipts or ID scans for private data before sharing the output.

Does it use the office-mcp server?

The frontmatter names it, but the instructions and code in SKILL.md never call it. The examples use PaddleOCR directly.

Information

updated Nov 2, 2026

Repository: claude-office-skills/skills
Folder: smart-ocr

claude-office-skills/skills skills.sh

Sources

Docs, issue trackers and discussions this page is built from.

Show all 2

smart-ocr SKILL.md · github.com
claude-office-skills README · github.com