Korean office documents (.hwp / .hwpx) are everywhere in Korean government, enterprise, and legal workflows. Until now, Claude, ChatGPT, and Cursor couldn't read them natively — a real blocker for anyone building AI systems for Korean organizations.
We (TreeSoop) just released hwp-mcp, an open source MCP server that fixes this.
📚 Full writeup: https://treesoop.com/blog/hwp-mcp-korean-document-ai-claude-2026
🔧 GitHub: https://github.com/treesoop/hwp-mcp
What it does
hwp-mcp exposes these tools via the Model Context Protocol:
- Extract text from .hwp / .hwpx files
- Parse tables into structured data
- Pull out embedded images
- Find-and-replace within documents
- Fill template variables (name, company, date)
Works on macOS and Windows. No Hancom Office license required.
Install in one line
claude mcp add hwp-mcp -- uvx --from hwp-mcp hwp-mcp
Works with Claude Code, Claude Desktop, VS Code Copilot, Cursor — anywhere that supports MCP.
Why this matters for Korean AI adoption
If you're building RAG systems, internal search, or document automation for Korean companies, 60–80% of the source documents will be HWP. Before hwp-mcp the options were:
- Manual conversion (doesn't scale)
- Hancom API licensing (Windows-only, paid)
- Convert everything to Word org-wide (non-starter)
Now you just install the MCP and Claude reads HWP natively.
What we're using it for
TreeSoop uses hwp-mcp in:
- Corporate RAG chatbots ingesting HWP knowledge bases
- Government RFP automation (RFPs are distributed as HWP)
- Legal contract review (Korean law firm contracts = HWP)
- Meeting-note template auto-fill
MIT licensed
Commercial use is fine. Contributions welcome.
- GitHub: https://github.com/treesoop/hwp-mcp
- Docs: included in repo README
TreeSoop is an AI-Native dev agency from Korea. POSTECH/KAIST team. We're building production AI agents, RAG systems, and MCP tools. More OSS: ai-news-mcp, whisper_transcription, claude2codex.
Top comments (0)