Skip to content
Get packs

Guide · Agent skill

How to use the LiteParse document parser agent skill

Updated · written with AI from the project's own documentation and checked against our catalogue data

LiteParse is an agent skill that lets an AI agent read PDFs, Word and Excel files, and images on your own machine, pull out their text, and record where each piece sits on the page. This guide explains what an agent skill is, how to install the liteparse skill, how to ask for it in plain language, and what to watch for around requirements, network use and safety. By the end you will be able to set it up and run typical parsing, OCR and batch jobs.

LiteParse Local Document ParserK-Dense-AI/scientific-agent-skills · MIT

Quick start

  1. 1Download and unzip. Download the skill ZIP from this page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder.
  2. 2Or upload in the Claude apps. In the Claude apps, open Settings, go to Capabilities, and upload the skill under Skills.
  3. 3Install LiteParse. Run uv pip install "liteparse==2.15.0" to get the Python bindings and the lit CLI.
  4. 4Install LibreOffice if needed. Add LibreOffice only if you will parse Word, Excel, PowerPoint or OpenDocument files.
  5. 5Verify the install. Run lit --help and check the liteparse version in Python.
  6. 6Ask for what you need. Describe the file and the output, for example text, JSON with bounding boxes, OCR or page PNGs.

What is an agent skill and what does LiteParse let the agent do?

An agent skill is a folder of instructions and helper files that an AI agent loads when a task matches the skill's description. The liteparse skill is built around the LiteParse open-source document parser, which works locally and returns layout-aware text.

With it, the agent can extract text from PDFs, DOCX, Office files and images; run OCR on scanned pages; produce structured JSON with bounding boxes or Markdown; render pages to PNG so a multimodal agent can see figures; and batch-process a folder of documents.

  • Layout-preserved text, structured JSON with bounding boxes, or heuristic Markdown
  • OCR on scans, using bundled Tesseract or an optional user-run HTTP OCR server
  • Page subsets and password-protected PDFs
  • Page screenshots as PNG files
  • Batch parsing of a directory

Install the LiteParse skill

Download the skill ZIP from this page and unzip it into one of these locations. For a personal skill available across projects, use ~/.claude/skills/. For a single project, use the .claude/skills/ folder inside that project.

Alternatively, in the Claude apps, open Settings, go to Capabilities, and upload the skill under Skills.

The skill also expects the LiteParse Python package and CLI. The source gives this install command, pinned to the tested version:

  • Personal: ~/.claude/skills/liteparse/
  • Project: .claude/skills/liteparse/ inside your project folder
  • Claude apps: Settings → Capabilities → Skills
uv pip install "liteparse==2.15.0"

Check that the tools work

The install provides the Python bindings and the lit CLI. Run these two checks to confirm both are available.

  • Install LibreOffice separately if you need Word, Excel, PowerPoint or OpenDocument inputs. Images such as PNG, JPEG, TIFF and WebP convert natively without it.
lit --help python -c "import liteparse; print(liteparse.__version__)"

Trigger the skill with natural requests

You usually do not need to name the skill. The agent reads the skill description and loads it when your request involves document text extraction. Ask for the outcome you want and name the file or folder.

Useful phrasings include asking for text from a PDF, OCR on a scanned report, a JSON file with bounding boxes for RAG, screenshots of specific pages so the agent can read a chart, or a batch conversion of a folder of papers.

  • "Extract the text from paper.pdf and tell me the main findings."
  • "Run OCR on scan_0042.pdf in French."
  • "Parse every PDF in ./papers to JSON with bounding boxes, including subfolders."
  • "Render pages 1, 3 and 5 of report.pdf as PNG so you can read the figures."

Common commands and workflows

The skill uses the lit command for quick jobs. Plain text is the default output. Add --format json for coordinates, and -o to write to a file. Use --no-ocr on text-native PDFs for speed, and --target-pages to limit the pages.

  • Parse specific pages and disable OCR for speed on digital PDFs
  • Batch-parse a folder recursively, restricted to .pdf files
lit parse paper.pdf --format json -o paper.json lit parse paper.pdf --no-ocr lit parse long_paper.pdf --target-pages "1-5,10" lit screenshot document.pdf --target-pages "1,3,5" -o ./screenshots lit batch-parse ./papers ./parsed --extension .pdf --no-ocr

Safety, network and what the skill needs

The skill runs scripts and uses the network. No cloud account is needed for normal local parsing. However, missing Tesseract language data is downloaded from GitHub the first time it is required, and an HTTP OCR server you configure receives document images. The skill is not a network-free tool on first run.

Use an offline setup by pre-populating every needed .traineddata file and setting TESSDATA_PREFIX or passing --tessdata-path. Do not send sensitive documents to an HTTP OCR server you do not control. The optional server needs its own authentication, which you supply through the server headers setting.

  • Requires Python 3.10 or newer and liteparse 2.15.0
  • LibreOffice is needed for Word, Excel, PowerPoint and OpenDocument inputs
  • Network access is needed for the first OCR language download and for any HTTP OCR server
  • The batch wrapper script writes outputs that differ from the native CLI JSON schema

Tips for reliable results

Raise the DPI for poor scans, since higher DPI improves OCR but runs slower. Set the OCR language with Tesseract codes such as fra or deu. For batch jobs, the native batch-parse command names outputs paper.json, so files with the same stem in one folder can collide; restrict the extension or use the bundled wrapper, which keeps suffixes such as paper.pdf.json.

Combine JSON parsing with screenshots when an agent needs both coordinates and images for the same pages. Check complex layouts visually, because Markdown output is heuristic.

  • Use --dpi 300 or dpi=300 for small or faint text, and expect slower runs
  • Pick a different tool for dense tables or handwriting, such as LlamaParse, which is a separate cloud service

What people use it for

  • Ingesting a folder of research papers and supplementary PDFs into a local RAG pipeline
  • Building citation features that highlight source regions using bounding boxes
  • Converting scanned contracts or reports to searchable text on your own machine
  • Rendering chart-heavy PDF pages to PNG so a multimodal agent can read figures

Frequently asked questions

Is LiteParse free to use?

The LiteParse skill is published under the MIT licence on the scientific-agent-skills repository. The source does not list any price, and no cloud account is needed for local parsing. Check the licence file in the repository for the exact terms.

What does LiteParse require?

It requires Python 3.10 or newer and liteparse 2.15.0 installed with pip or uv. LibreOffice is needed for Word, Excel, PowerPoint and OpenDocument files, and internet access is needed the first time missing Tesseract language data is downloaded.

Does LiteParse send my documents to the cloud?

Local parsing runs on your machine. However, an HTTP OCR server you configure receives document images, so avoid it for sensitive files unless you control that server. Missing OCR language data is downloaded from GitHub.

How do I run OCR offline?

Pre-populate every .traineddata file you need, then set TESSDATA_PREFIX or pass --tessdata-path. A directory setting alone does not stop downloads, so the files must already be present.

How is LiteParse different from LlamaParse?

LiteParse runs locally and returns text, bounding boxes, Markdown and PNG pages. LlamaParse is a separate cloud service that requires its own sign-up and is suggested for dense tables, handwriting or production cloud pipelines.

Why does my batch output have colliding file names?

The native lit batch-parse command writes paper.json for each input, so same-named files in one folder can overwrite each other. Restrict the input extension or use the bundled scripts/batch_parse_dir.py wrapper, which keeps suffixes like paper.pdf.json.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

LiteParse suits researchers and developers who want local text extraction with bounding boxes, OCR and batch ingestion, and who can accept first-run downloads for OCR language data. Pick another tool if you need to merge or edit PDFs, process dense tables or handwriting, or send documents to a cloud service.

Download the skill