LibreOffice MCP 2.0

An MCP server that lets AI assistants (Claude Desktop, Claude
Code, and any other MCP client) create, read and edit documents with LibreOffice (Writer, Calc, Impress)
— no GUI, no copy-paste.

Ask: "Write a one-page project brief as brief.docx with a title, three sections, a risks
table and page numbers in the footer, then export it to PDF." — and the assistant builds it.

How it works (read this first)

• The Docker container is the MCP server. LibreOffice, its Python bridge and fonts are all
  inside the image. You build the image once.
• Your AI app (Claude Desktop, Claude Code…) only needs to be told how to reach it
  (claude mcp add … or a few lines of JSON). It does not install anything else.
• The AI can only see one folder on your computer: the workspace you mount into the container
  as /data. See The workspace folder.

Quick start

Step 1 — build the image (once)

Needs Docker. The first build takes a few
minutes (it downloads LibreOffice); later builds are cached.

git clone https://github.com/harshithb3304/libre-office-mcp.git
cd libre-office-mcp
docker build -t libreoffice-mcp .
docker run --rm libreoffice-mcp doctor      # optional check: should print [ok] twice

Step 2 — connect your AI app (pick one option)

Option A — the AI app starts the container itself (simplest; nothing stays running).
The part after -v is <a folder on your computer>:/data.

claude mcp add libreoffice -- docker run -i --rm -v /path/to/my-workspace:/data libreoffice-mcp

(Windows: %APPDATA%\Claude\, macOS: ~/Library/Application Support/Claude/)

{
  "mcpServers": {
    "libreoffice": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "-v", "C:\\Users\\me\\my-workspace:/data", "libreoffice-mcp"]
    }
  }
}

Restart Claude Desktop. (On Windows, double every backslash in JSON.)

Option B — keep the server running in the background (docker compose; instant responses, one
URL for any client). Start it, then just register its address:

docker compose up -d --build        # starts the server on http://localhost:8000/mcp
claude mcp add --transport http libreoffice http://localhost:8000/mcp

Compose uses the folder ./workspace next to docker-compose.yml (created automatically). To
use a different folder: WORKSPACE=/path/to/my-workspace docker compose up -d
(or put WORKSPACE=... in a .env file next to the compose file).

There is no authentication; the port is published on 127.0.0.1 only, so just your
computer can reach it. Read docs/DOCKER.md before exposing it further.

Step 3 — try it

Ask your AI: "Create hello.docx with a heading and a three-item bullet list."
Then open hello.docx from your workspace folder.

Everyday commands

The workspace folder

The container cannot see your computer; it only sees the folder you mount. We call it the
workspace, and inside the container it appears as /data.

• Relative paths the AI uses (report.docx, clients/acme.docx) resolve to the workspace.
  It is not your terminal's current directory unless you mount it (-v "$(pwd):/data").
• Absolute paths work only inside /data; anything else is refused. Paths from your own
  computer (C:\Users\…) don't exist in the container, so just give file names.
• It lives on your disk, not in the image or container. Stopping or deleting the container
  does not touch it, and the image does not grow as you create documents.
• To work on an existing file, copy it into the workspace first.
• Close a file in other programs before the AI edits it. Some programs lock open files (the
  write is then refused); others keep showing the old content until you reload.

More detail and permissions on Linux hosts: docs/DOCKER.md.

Choosing which tools the AI gets

By default everything is enabled. More tools means more for the model to read, so you can
narrow it down by adding flags after the image name:

Example: claude mcp add libreoffice -- docker run -i --rm -v /path/ws:/data libreoffice-mcp --apps writer.
docker run --rm libreoffice-mcp tools --apps writer prints exactly which tools a set of flags
exposes. Every flag also has a LIBREOFFICE_MCP_* environment variable. The office_* core tools
(create, convert, info, batch edit…) stay available unless excluded.

What the AI can do

The full list with parameters is in docs/TOOLS.md (generated from the code).

Writer

• Read — text (paged), structure (every paragraph/table with an index), outline, find, paragraph formatting, styles, tables, images, comments, bookmarks, links, footnotes, page setup, headers/footers, tracked changes
• Write — paragraphs, headings, bullet/numbered lists, whole sections in one call (writer_add_content), page breaks, merge another file
• Format — bold/italic/colour/font/size on specific text, paragraph alignment & spacing, create/modify styles
• Tables — create, edit cells, add/delete rows & columns, merge cells, borders, shading, header rows, column widths
• Objects — images (embedded, resizable, alt text), hyperlinks, bookmarks, footnotes, review comments, table of contents
• Page — paper size, orientation, margins, header/footer text with “Page N of M”
• Review — turn change tracking on/off, accept or reject all changes

Any document — create, info, copy, convert/export (PDF, docx, odt, html, txt…), and
office_batch_edit (many edits, one load, one save, all-or-nothing).

Writer content is addressed by block index (0 = first paragraph or table): the AI calls
writer_get_structure to see the indexes, then edits by index — no fragile text matching.

Safety

• Your file is never half-written. Each request edits a private copy and replaces your file atomically at the end. A crash, a timeout or a failing step in a batch leaves the original untouched.
• No races. Requests run one at a time, so two simultaneous edits to the same file never interleave (the second waits).
• Macros never run, even in documents that contain them.
• The container only sees the workspace (/data); paths outside it, .. and symlink escapes are refused.
• No lock files are created next to your documents, and a stuck or cancelled request is killed cleanly (no leftover LibreOffice processes).
• There is no undo. Edits are saved immediately; keep backups of important files (or use office_copy_document first).

Coming from v1?

v2 is a rewrite; tool names changed to be consistent (office_*, writer_*):

The old libreoffice:{path} MCP resource is not ported yet.

Documentation

License

MIT