← Back to Skills Marketplace
939
Downloads
0
Stars
0
Active Installs
1
Versions
Install in OpenClaw
/install jpocr
Description
Japanese OCR via NDLOCR-Lite (National Diet Library). Trigger on 'OCR this image', '日文OCR', 'recognize Japanese text', or any request to extract text from Ja...
README (SKILL.md)
jpocr — Japanese OCR Skill
Local Japanese OCR powered by NDLOCR-Lite from Japan's National Diet Library. Runs on CPU (Apple Silicon / x86), no GPU or API key required.
Capabilities
| Target | Quality |
|---|---|
| Printed Japanese (活字) | Excellent |
| Vertical text (縦書き) | Excellent |
| English text | Good |
| Handwritten Japanese (手書き) | Experimental |
How to call
Run scripts/ocr-cli.sh from the skill root directory:
\x3CSKILL_ROOT>/scripts/ocr-cli.sh \x3Cimage_path> # → plain text to stdout
\x3CSKILL_ROOT>/scripts/ocr-cli.sh \x3Cimage_path> --json # → JSON with bounding boxes
\x3CSKILL_ROOT>/scripts/ocr-cli.sh \x3Cimage_path> --viz # → also saves visualization
\x3CSKILL_ROOT>/scripts/ocr-cli.sh \x3Cdir_path> # → batch all images in dir
Output formats
text (default): one line per detected text region.
json:
{
"contents": [[
{
"boundingBox": [[x1,y1],[x1,y2],[x2,y1],[x2,y2]],
"text": "recognized text",
"confidence": 0.95,
"isVertical": "true"
}
]],
"imginfo": { "img_width": 1920, "img_height": 1080 }
}
viz: saves viz_\x3Cfilename> bounding-box overlay image to the output directory.
Performance
- ~2-3 seconds per image on Apple Silicon (CPU)
- Formats: JPG, PNG, TIFF, JP2, BMP
- Charset: ~7000 characters (JIS kanji + kana + ASCII + Greek)
Tech stack
- Layout detection: DEIMv2 (ONNX)
- Text recognition: PARSeq cascade (30/50/100 char models, ONNX)
- Reading order: xy-cut algorithm
Usage Guidance
This package is incomplete and opaque. Before installing or running: (1) request or inspect the missing files (src/ocr.py, model assets, and any venv setup scripts) and any instructions for obtaining model weights; (2) verify where those models/code would be downloaded from (trusted GitHub release or official NDLOCR distribution) — avoid skills that fetch code from unknown servers; (3) note that the runner script silences stdout/stderr (>/dev/null 2>&1), which will hide errors and any network download messages—run it in a sandbox or with that redirection removed so you can see activity; (4) prefer a skill with a clear install spec or documented provenance for large model files. If the author cannot provide the missing files and a clear install provenance, treat this skill as untrusted.
Capability Analysis
Type: OpenClaw Skill
Name: jpocr
Version: 1.0.0
The `scripts/ocr-cli.sh` file contains a critical shell injection vulnerability. The `$SOURCE_ARG` variable, which is constructed using user-controlled input (`$INPUT`), is used unquoted in the execution of the Python script (`"$VENV" "$OCR" $SOURCE_ARG ...`). This allows an attacker to inject arbitrary shell commands by crafting the `<image_path>` argument, leading to potential Remote Code Execution (RCE). While the `SKILL.md` instructions themselves are not malicious prompt injection, the underlying script's vulnerability makes this skill suspicious.
Capability Assessment
Purpose & Capability
Name/description promise local OCR via NDLOCR-Lite, which is plausible, but the repository bundle lacks the referenced runtime (src/ocr.py) and virtualenv (.venv/bin/python). The skill therefore cannot perform its stated purpose as packaged; that mismatch is incoherent.
Instruction Scope
SKILL.md instructs running scripts/ocr-cli.sh which invokes a python program and writes output files. The README provides no installation steps, model files, or network/download behavior. The executor script redirects the OCR process's stdout/stderr to /dev/null, hiding errors or network activity—this opaque behavior is risky and the instructions are incomplete.
Install Mechanism
There is no install spec (low risk), but the script expects a prebuilt .venv and model/code under src/ which are not included. Absence of an install mechanism or documentation for obtaining models/code is a packaging/maintainability concern (it may require fetching large model artifacts or code from elsewhere).
Credentials
The skill requests no credentials and only optionally uses JPOCR_OUTPUT to set output directory. No secret or unrelated environment variables are required. However, the use of a hidden .venv path and optional env var should be documented.
Persistence & Privilege
Flags are default (not always-on); the skill does not request persistent system-wide presence or modify other skill configs. It appears to run as a one-off command when invoked.
How to Use
- Make sure OpenClaw is installed (local or Docker)
- Run the install command in chat:
/install jpocr - After installation, invoke the skill by name or use
/jpocr - Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.0.0
- Initial release of jpocr, a local Japanese OCR skill powered by NDLOCR-Lite from Japan's National Diet Library.
- Supports high-quality printed and vertical Japanese text recognition; works for English text as well.
- Provides batch processing and multiple output formats: plain text, JSON with bounding boxes, and visualization images.
- Experimental support for handwritten Japanese OCR.
- Runs on CPU (Apple Silicon/x86); no GPU or API key required.
Metadata
Frequently Asked Questions
What is jpocr?
Japanese OCR via NDLOCR-Lite (National Diet Library). Trigger on 'OCR this image', '日文OCR', 'recognize Japanese text', or any request to extract text from Ja... It is an AI Agent Skill for Claude Code / OpenClaw, with 939 downloads so far.
How do I install jpocr?
Run "/install jpocr" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.
Is jpocr free?
Yes, jpocr is completely free (open-source). You can download, install and use it at no cost.
Which platforms does jpocr support?
jpocr is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).
Who created jpocr?
It is built and maintained by Wayne (@realwaynesun); the current version is v1.0.0.
More Skills