kordoc

What it does

kordoc turns Korean documents (HWP 3.x and 5.x, HWPX, HWPML, PDF, XLS/XLSX, DOCX, PPTX and PNG/JPG/WebP images) into Markdown and structured data. It comes as a JavaScript library, a command-line tool and an MCP server. It reads Hancom files itself, so Hancom Office does not need to be installed.

Start the MCP server with `npx -y kordoc mcp` and you get 22 tools (counted in src/mcp.ts and its six tool modules). Besides parsing whole documents, page ranges, single tables, metadata or RAG chunks, it can diff two documents, read and fill form fields, and write edited Markdown back into the original HWPX or HWP. It also generates HWPX from Markdown, places seals, adds memos, masks personal data, and renders pages to images, SVG, HTML or PDF so the output can be checked. Newer additions are `convert_to_hwpx`, which turns an HWP file into HWPX, and agency packs: `build_agency_pack` learns one office's house style from a few of its HWPX files and saves it as a .kpack that `generate_document` can then write with. `lint_document` checks date, amount and loanword notation on every run, and can also check bullet style, AI-style phrasing, report structure, figures and budget sums when asked.

`npx -y kordoc setup` registers the server in installed clients such as Claude Desktop, Claude Code, Cursor and Codex. OCR for scanned pages runs on the local CPU without an API key. According to src/shared/offline.ts, the only outside traffic is the first download of OCR and formula models from huggingface.co and the CLI's `watch --webhook` URL. KORDOC_OFFLINE=1 blocks both before they are tried, and KORDOC_ROOT confines file access to one directory (no limit when unset).

What it needs

Limits

• Size cap lint_document checks at most 1,000,000 characters of text and returns at most 500 findings per list; longer documents must be split
• Not covered Form filling, patching, generation, seals, memos and rendering work on HWPX/HWP; other formats are read as Markdown only
• Not covered convert_to_hwpx carries over only text and tables from HWP 3.x files, and does not convert distribution (read-only) documents
• Limit convert_to_hwpx will not overwrite an existing .hwpx at its default path; give output_path to replace it (src/mcp/tools-convert.ts)
• Limit build_agency_pack takes HWPX samples only (1 to 50); HWP samples must be saved as HWPX in Hancom first
• Limit OCR and formula OCR models (about 18 MB and 155 MB) are downloaded on first use unless KORDOC_OFFLINE is set and models are side-loaded
• Platform On Windows, Claude Desktop may need the command written as cmd /c npx -y kordoc mcp

Who it suits

Suits people who work with Korean public-sector or office documents in HWP/HWPX and want an assistant to read, check, fill or draft them on their own machine, without Hancom Office and without sending files to a hosted service. Because it writes files and has no folder limit by default, set KORDOC_ROOT (and KORDOC_OFFLINE on closed networks) before pointing it at a shared disk. If you only need plain-text extraction from common Office formats, a general converter may be lighter; kordoc's value is the Hancom formats and Korean official-document tooling.