borg-mcp
A read-only MCP server for BorgBackup that lets AI agents query backup status, repository info, archive listings, and prune previews without modifying or destroying backups.
README
<p align="center"> <img src="assets/logo.svg" alt="borg-mcp" width="480"> </p>
A read-only MCP server for BorgBackup operational status.
borg-mcp lets AI agents (Claude, and other MCP clients) answer questions like "is my backup fresh?", "how big is the repo?", "what would prune remove?" — without any ability to modify or destroy backups.
It is a companion project to borg, born from borg#9955. It is intentionally not part of borg core: the safety boundary for untrusted AI agents lives in this small, auditable project.
Status: pre-alpha, under development. Requires borg2 (no borg 1.x support).
Security model
The MCP client (the agent) is treated as untrusted input — it may be prompt-injected or simply wrong. Therefore:
- Command allowlist, not passthrough. Only a fixed set of borg subcommands with fixed argument templates is ever executed. No agent-supplied flags, no shell.
- Repository allowlist. Repositories are configured server-side with
aliases; the agent refers to
"home", never to a raw path or URL. - No secrets over MCP. Passphrases come from configured passcommands on the server side and never appear in tool results or logs.
- Sanitized, size-capped output. Raw file listings are off by default (explicit config opt-in); all listings are paginated. borg's stderr is sanitized before an error reaches the agent: tracebacks are withheld (logged server-side instead), the passcommand line is redacted, and control characters are stripped.
- Read-only by construction — and by enforcement. borg-mcp never runs
prune/delete/compact/repair/restore. For defense in depth, run it against
repositories accessed via an SSH key that is restricted server-side to
read-only
borg serve, and run borg-mcp itself as an unprivileged user. - Audit log. Every tool invocation is logged.
Never in scope: executing destructive operations, key export, passphrase handling over MCP, arbitrary borg commands.
Requirements
- Python >= 3.11
- borg2, installed from the
master branch (borg-mcp drives the
borgCLI; it does not import borg internals)
Installation
There is no PyPI release; install from git:
pip install git+https://github.com/borgbackup/borg-mcp.git
or from a checkout: pip install .
Configuration
TOML, at ~/.config/borg-mcp/config.toml (or /etc/borg-mcp.toml).
The config file holds all security-relevant settings and must not be
writable by the agent's account — there are deliberately no CLI overrides.
[server]
allow_file_listing = false # raw archive contents listing, off by default
timeout = 300
max_items = 1000 # pagination cap for listings
[repos.home]
location = "ssh://backup@host/./home"
description = "workstation home dirs, nightly"
passcommand = "cat /path/to/passphrase-file"
The passcommand must read the passphrase from somewhere safe: a file only
readable by the borg-mcp user, or a keyring/secret-service CLI. Never put
the passphrase inline (as in passcommand = "echo secret"): the command
line is visible in ps output while it runs.
Usage
borg-mcp serve [--config PATH] # run the stdio MCP server
borg-mcp check [--config PATH] [ALIAS ...]
# validate config; with aliases, test
# repository access
MCP client configuration (e.g. Claude Code .mcp.json):
{
"mcpServers": {
"borg": {
"command": "borg-mcp",
"args": ["serve", "--config", "/path/to/config.toml"]
}
}
}
Tools
| tool | answers |
|---|---|
list_repositories |
which repositories can I ask about? |
repo_info |
repository ID, size, encryption mode |
list_archives |
which archives exist? (paginated, filterable) |
archive_info |
stats, duration, hostname for one archive |
latest_archive |
is my backup fresh? |
prune_preview |
what would a prune with these keep rules remove? |
prune_preview never deletes anything: --dry-run is part of the command
template and cannot be influenced by the client, and the test suite asserts
that every prune command borg-mcp can construct contains it.
These tools exist only when allow_file_listing = true, because they
disclose the names of the backed up files:
| tool | answers |
|---|---|
list_archive_contents |
which files are in this archive? |
diff_archives |
which files changed between two archives? |
License
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。