opencode-parser

An opencode plugin that parses any file into structured text the LLM can work with.

This plugin supports both OpenCode V1 and V2 from the same package. V1 calls server(), V2 calls setup() (requires OpenCode V1 >= 1.18.29 for the object entrypoint).

Opencode-parser.mp4

Install

OpenCode V2

{
  "plugins": ["opencode-parser"]
}

OpenCode V1

{
  "plugin": ["opencode-parser"]
}

Or install via CLI: opencode plugin opencode-parser -g

Or copy src/ into .opencode/tools/ for a local zero-config setup.

Supported formats

Format	Extensions	Extracted
PDF	.pdf	Text, metadata, pages
Word	.docx	Text, tables, metadata
Excel	.xlsx, .xls, .csv, .tsv	Text, tables, sheet names
PowerPoint	.pptx, .ppt	Slide text, speaker notes
Images	.png, .jpg, .jpeg, .webp, .gif, .bmp, .tiff	OCR text (opt-in)
EPUB	.epub	Full text with heading structure
HTML	.html, .htm	Body text, headings
XML	.xml	Stripped text content
Markdown	.md	Raw text
Jupyter	.ipynb	Code, markdown, outputs
ZIP	.zip	File listing with sizes
Archives	.rar, .7z, .tar, .gz	Listing (extraction notes)
Plain text	.txt, .json, .yaml, .toml, .ini	Raw content

Usage

Parse @report.pdf and give me a summary

parse the spreadsheet at @data.xlsx but only the first 3 sheets

parse @report.pdf and save the full output

Options

Option	Default	Description
filePath	—	Path to the file (required)
maxChars	50000	Limit output chars (-1 for unlimited). Pass -1 to get the full document.
extractTables	true	Extract tables from docs/spreadsheets
extractImages	false	Enable OCR for images
ocrLang	"eng"	OCR language for tesseract.js (e.g. "eng", "fra", "ara")
maxPages	varies	Limit pages/slides/sheets processed
save	false	Save the full parsed output as a .md file alongside the original (no truncation)
outputPath	—	Custom path for the Markdown export (overrides save path)

How it works

1. File is verified by magic bytes, not just extension
2. Type detection dispatches to the right parser
3. Metadata is extracted (author, pages, sheet count, etc.)
4. Tables become readable markdown
5. Large content is truncated gracefully with a note to the LLM

All 15+ format handlers return the same output structure, so the LLM gets consistent results regardless of file type.

Development

npm install
npm run typecheck

License

MIT