edit2docs

AI-agent-native document engine — DOCX · XLSX · PPTX. Korean-first.

한국어 README · Sister project of edit2ppt (the PPTX pipeline is inherited from it) · Built on ppt-master (MIT)

edit2docs generates complete Office documents from a one-line intent and chat-edits existing files — Word reports, Excel workbooks and PowerPoint decks — always producing natively editable OOXML. One engine, four surfaces:

pip install edit2docs              # library + agent tools + local MCP
pip install "edit2docs[server]"    # + the hosted multi-tenant service

The five verbs

Every surface exposes the same five format-dispatched verbs — the file extension picks the engine:

1 · Python library

from edit2docs import generate_doc, edit_doc, preview_doc, analyze_doc, set_doc_text

generate_doc("3분기 실적 보고서", output="report.docx")
generate_doc("분기별 매출 정리", output="sales.xlsx", sources=["raw.pdf"])
generate_doc("Q3 영업 결과 임원 보고", output="deck.pptx", template="brand.pptx",
             deck_mode="template_restyle")

r = edit_doc("report.docx", "진행 사항 섹션에 배포 완료 항목을 추가해줘")
print(r.reply, r.operations)

info = analyze_doc("sales.xlsx")               # sheets + sample rows
set_doc_text("sales.xlsx", [{"sheet": "매출", "cell": "B3", "value": 142}])

BYOK: api_key=... or ANTHROPIC_API_KEY. The deterministic verbs need no key. Async variants: async_generate_doc, async_edit_doc.

2 · Agent tools (function calling)

from edit2docs.agent_tools import ANTHROPIC_TOOLS, run_tool

msg = client.messages.create(model="claude-opus-4-7", tools=ANTHROPIC_TOOLS, ...)
for block in msg.content:
    if block.type == "tool_use":
        result = run_tool(block.name, block.input)

3 · Local MCP server (zero infra)

// Claude Desktop / Claude Code / Cursor
{ "mcpServers": { "edit2docs": {
    "command": "edit2docs-mcp",
    "env": { "ANTHROPIC_API_KEY": "sk-ant-..." }
} } }

4 · Hosted service

pip install "edit2docs[server]" then edit2docs serve — the FastAPI service (REST + SSE jobs + hosted MCP) inherited from edit2ppt: multi-tenant assets, job queue, S3-compatible storage, PPTX studio endpoints.

How each format works

• DOCX — the writer LLM emits a constrained markdown document; a deterministic renderer (python-docx) turns it into styled Word. Edits are paragraph-addressed operations (replace / insert_after / delete, table cells by row/col) so untouched paragraphs keep their formatting. The hosted preview is a native, addressable HTML rendering: every paragraph carries data-e2d-para and every table cell data-e2d-table/data-e2d-cell — the same addresses the outline and the live-edit op stream use — plus real merged cells, alignment, colors, images, footnotes and page breaks (mirroring the PPTX canvas's data-e2p-* tags).
• XLSX — the designer LLM emits a YAML sheet spec (sheets / headers / rows / number formats, formulas allowed); openpyxl renders styled sheets. Edits are set_cell / append_rows / add_sheet with staleness guards. The hosted preview renders a spreadsheet-style grid (column letters, row numbers, merged ranges, cached formula results) with every cell stamped data-e2d-cell="B3" — exactly the address set_cell takes.
• PPTX — the full multi-stage deck pipeline inherited from edit2ppt: strategist → per-page SVG → native DrawingML, user-PPTX templates (restyle/extend), chat-edit with slide recompose, per-paragraph text edits incl. table cells.

Every LLM planner follows the same contract: fenced reply + edit_plan blocks, one retry with a format reminder, and an honest reply (instead of a silent no-op) when planning fails. Korean-native throughout: Hangul-aware widths, lang="ko-KR" OOXML runs, bilingual errors.

License

MIT. PPTX core forked from ppt-master (MIT) via edit2ppt.