python-umcp-calibre
Read-only MCP server that runs inside Calibre as a plugin, exposing the live Calibre library for inspection, search, metadata, and duplicate detection via Streamable HTTP.
README
python-umcp-calibre
Calibre is quite particular about who touches its active library--and rightly so, since the GUI keeps database and filesystem state in memory. I wrote python-umcp-calibre to expose that live library to MCP clients without putting a second process in charge of metadata.db.
The server runs inside Calibre as an Interface Action plugin and uses umcp for Streamable HTTP. There is no sidecar: an MCP client connects directly to /mcp, and every library call reaches gui.current_db through a serialised worker.
What It Does
Clients can inspect the active library, search for books, fetch metadata, find probable duplicates, check an existing authenticated content server and inspect bridge audit records. capabilities_readonly gives agents a compact starting point, while describe_tool_readonly expands one tool at a time.
Mutations stay hidden until the Calibre UI has both saved a token and enabled them. Once that gate is open, the plugin can update metadata and formats, replace covers, import books, convert formats, export through Calibre's save-to-disk engine, copy or move books between allowlisted libraries, submit an existing format to a configured e-mail recipient, merge duplicates conservatively and move confirmed deletions to Calibre trash.
The slower paths appear in Calibre's own Jobs UI. Conversion uses a worker-process job; import, export, copy/move and e-mail use ThreadedJob. Move is deliberately fussy: preview -> copy -> metadata and format-hash verification -> source trash. Any uncertainty leaves the source in place.
A few boundaries are intentional. Permanent deletion, arbitrary e-mail recipients, implicit e-mail conversion, unauthenticated or temporary public links, and device actions are not exposed. Calibre 9.11 has no safe scoped temporary-link API, and device operations depend on live GUI/device state that the bridge cannot treat as a stable library mutation.
The plugin also exposes GET /health, plus Start, Status and Stop actions in the Calibre GUI. Non-loopback binds require a bearer token.
This release is source-contract and runtime tested against exactly Calibre 9.11.0 (v9.11.0, commit b23dfb5d). The plugin declares 9.11.0 as its minimum; later Calibre releases may still load the read-only surface, but mutation discovery fails closed until that exact runtime has been audited again.
Building It
PYTHONPATH=.:src python3 -W error::ResourceWarning -m unittest discover -s tests -v
sh plugins/build-plugin.sh
The build produces plugins/calibre-umcp-plugin.zip. It copies umcp.py and umcp_shared.py from src/calibre_umcp into the archive, so the plugin uses the same µMCP runtime as the rest of the repository rather than carrying a second protocol implementation.
Installing It
Install the ZIP with Calibre's plugin utility:
calibre-customize -a plugins/calibre-umcp-plugin.zip
linuxserver/calibre profiles are normally owned by abc, so install as that user inside the container:
s6-setuidgid abc calibre-customize -a plugins/calibre-umcp-plugin.zip
Restart or reload Calibre after replacing the plugin. It starts MCP automatically once the active library is available; the µMCP Bridge menu still provides Status, Stop and Start controls.
Connecting
The default bind is local to the Calibre process:
CALIBRE_UMCP_BRIDGE_HOST=127.0.0.1
CALIBRE_UMCP_PORT=9000
To reach it across a container or LAN network, bind explicitly and set a long random token:
CALIBRE_UMCP_BRIDGE_HOST=0.0.0.0
CALIBRE_UMCP_PORT=9000
CALIBRE_UMCP_BRIDGE_TOKEN=<long-random-token>
The MCP URL is http://<calibre-host>:9000/mcp. Authenticated clients send Authorization: Bearer <token>.
The architecture notes explain the process boundary, the design notes cover the implementation choices, and the plugin README has the container-oriented installation details.
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。