sokgi-mcp 0.2.1

Sokgi — read and write Korean business documents (HWP/HWPX/DOCX/XLSX/PPTX) from your AI agent

속기 (Sokgi) — MCP server

Korean business documents your AI agent can actually read.

An LLM will happily write a parser for .hwpx — and on real Korean government documents it gets 27.9% of tables wrong, silently. One unhandled merged cell shifts a whole row, so "담당 부서" comes back as "담당자". Nothing errors.

Sokgi restores tables properly — on your machine — and hands your agent the fields it asked for.

Tools

Tool	What it does
read_document(paths)	Parse locally → text + tables as 0-based grids, merged cells expanded
extract_documents(paths, fields)	Pull named fields out of a batch of documents
fill_document(template, values, out_path)	Fill a HWPX template, save a new file

Formats: HWPX / DOCX / XLSX / PPTX (text + tables) and HWP 5.0 (text). No Hancom Office required — reading uses a pure-Python OLE parser, writing produces HWPX (ZIP + XML). PDF is not supported; asking for one returns an explicit error, not an empty document.

Install

claude mcp add sokgi --env SOKGI_API_KEY=<key> -- uvx sokgi-mcp

Get a free key at https://sokgi.14dimension.com/console.

Why it abstains

A field that is not in the document comes back null and listed in missing — never invented. You delegate precisely because you cannot re-check every file, so a confident guess is worse than useless. Likewise fill_document refuses to save a document with unreplaced placeholders or a table too narrow for its text.

Where your documents go. Documents are uploaded to the Sokgi API, processed there, and deleted right after — they are not stored. Results come back to this machine, and fill_document writes the finished file here.

If your documents may not leave your network at all, that is what the on-premise install is for: it runs the whole pipeline inside your own network (SOKGI_LOCAL_PARSE=1). Ask us.

---

Part of the 명철 (MyungChul) series · https://ai.14dimension.com · [e-mail]