libre-mcp
MCP server for controlling LibreOffice (Writer and Calc) via the UNO API, enabling document creation, editing, saving, and export from AI clients.
README
libre-mcp
A stdio MCP server for controlling LibreOffice (Writer and Calc) from Claude Code and other MCP clients, over the UNO automation API. macOS and Linux.
Architecture
Two processes, both Python:
Full docs: libre-mcp.krondor.org.
Claude Code ──stdio JSON-RPC──▶ libre-mcp server (uv, Python >=3.12, mcp SDK)
│ spawns + supervises
├─▶ headless soffice (isolated profile + named pipe)
└─▶ uno_worker.py (LibreOffice's Python, stdlib-only)
└─ newline-JSON RPC over stdio
- The server owns the MCP stdio channel, defines the tools, and supervises
the other two processes. It runs on a normal uv-managed Python ≥3.12 with the
mcpSDK. - A headless soffice is launched with an isolated
UserInstallationprofile and a unique named pipe (not a TCP port), so concurrent instances never race over a shared port — important for parallel git worktrees. - The UNO worker (
src/office/uno_worker.py) runs under LibreOffice's Python — the only interpreter that can reliablyimport uno. It is intentionally stdlib-only, so there is nothing to install alongside LibreOffice. (On Debian/Ubuntu it falls back to the systempython3provided bypython3-uno.)
Instances are disposable: if the URP bridge is disposed (soffice died), the worker reports a fatal error and the session is rebuilt, retrying the call once.
Install
curl -fsSL https://raw.githubusercontent.com/krondor-corp/libre-mcp/main/install.sh | bash
claude mcp add libre -- libre-mcp
Downloads a prebuilt, self-contained binary (no Python needed on the host).
Update in place with libre-mcp update.
Requirements
- LibreOffice installed:
- macOS:
/Applications/LibreOffice.app - Debian/Ubuntu:
apt install libreoffice-writer libreoffice-calc python3-uno - TDF/opt Linux builds bundle their own Python and need nothing extra.
- macOS:
Paths are auto-discovered; override with LIBRE_MCP_SOFFICE_PATH /
LIBRE_MCP_PYTHON_PATH if needed. (uv is only needed for development.)
Develop
make sync # uv sync (dev deps)
make check # fmt-check + lint + types + test
make inspect # run the MCP Inspector against the server
make test runs unit tests always; the end-to-end tests run only when
LibreOffice is found, otherwise they skip.
Run it locally (dev client)
dev/client.py spawns the server exactly as an MCP host would, does the
initialize handshake, and drives tools — no registration needed.
make dev-list # list the tools
make dev-demo # scripted writer -> PDF round-trip
For ad-hoc multi-step sessions, pipe TOOL {json-args} lines into run. A
single run keeps one server (and one LibreOffice) alive for the whole batch,
so doc_ids persist across calls (first created doc is doc-1, next doc-2, …):
uv run python dev/client.py run <<'EOF'
create_document {"kind": "calc"}
set_cells {"doc_id": "doc-1", "cells": [{"cell": "A1", "value": 21}, {"cell": "A2", "formula": "=A1*2"}]}
read_cells {"doc_id": "doc-1", "range": "A1:A2"}
EOF
Register with Claude Code
A project-scoped .mcp.json is committed with a relative command
(./bin/run.sh), so a fresh checkout / worktree self-registers and runs its own
copy. Or register explicitly:
claude mcp add --transport stdio libre -- $(pwd)/bin/run.sh
Parallel worktrees
The design is worktree-safe by construction:
- Unique named pipe per instance — no shared TCP port to race over, so two worktrees (or two sessions) starting at the same moment never collide.
- cwd-relative profiles —
.lo_profiles/instance-<pipe>lives under each worktree's own directory, fully isolated and gitignored. - Clean reaping — the server tears down its soffice + worker when the client disconnects (lifespan hook), so worktrees don't leak LibreOffice processes.
- Relative
.mcp.json— each worktree registers/runs its ownbin/run.sh.
Tools
| Tool | Purpose |
|---|---|
create_document(kind) |
New writer/calc/impress/draw doc → doc_id |
open_document(path) |
Open a file → doc_id |
list_documents() / document_info(doc_id) |
Inspect open docs |
save_document(doc_id, path?, format?) |
Save in place or to a path |
export_document(doc_id, path, format?) |
Export (pdf, docx, xlsx, csv, …) |
close_document(doc_id) |
Close a doc |
get_text(doc_id) |
Read a Writer doc's text |
insert_text(doc_id, text, paragraph_break?) |
Append text |
find_and_replace(doc_id, search, replace, regex?) |
Replace text |
set_cells(doc_id, cells, sheet?) |
Write Calc cells/formulas |
read_cells(doc_id, range, sheet?) |
Read a Calc range |
Configuration
| Env var | Default | Meaning |
|---|---|---|
LIBRE_MCP_SOFFICE_PATH |
auto | soffice binary override |
LIBRE_MCP_PYTHON_PATH |
auto | bundled python override |
LIBRE_MCP_PROFILE_DIR |
./.lo_profiles |
per-instance profile root |
LIBRE_MCP_KEEP_PROFILE |
false |
keep profiles for debugging |
LIBRE_MCP_STARTUP_TIMEOUT |
30 |
seconds to wait for soffice |
LIBRE_MCP_LOG_LEVEL |
INFO |
stderr log level |
UNO transport uses a unique named pipe per instance (not a TCP port), so nothing needs configuring and concurrent instances never collide.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。