SSH MCP Server
Enables Claude to execute commands on remote servers via SSH, with support for SSH config, private keys, password auth, and agent forwarding.
README
SSH MCP Server
An MCP (Model Context Protocol) server that lets Claude execute commands on remote servers over SSH. Supports ~/.ssh/config, private keys, password auth, and SSH agent forwarding.
Tools
| Tool | Description |
|---|---|
ssh_connect |
Open an SSH connection to a remote host |
ssh_execute |
Run a command on a connected host |
ssh_disconnect |
Close a connection |
ssh_list_connections |
Show all connections and their status |
Setup
git clone <repo-url> && cd ssh-mcp-server
npm install
npm run build
Claude Code
claude mcp add ssh node /absolute/path/to/ssh-mcp-server/build/index.js
Claude Desktop
Add to claude_desktop_config.json:
{
"mcpServers": {
"ssh": {
"command": "node",
"args": ["/absolute/path/to/ssh-mcp-server/build/index.js"]
}
}
}
Authentication
Auth is resolved in this order — the first match wins:
- Explicit password —
passwordparam - Explicit private key —
privateKeyPathparam (supportspassphrase) - SSH config IdentityFile — reads
~/.ssh/config, tries each key in order - SSH agent — uses
SSH_AUTH_SOCK(default when nothing else is specified)
Using ~/.ssh/config
The server reads ~/.ssh/config on every connect call, so edits are picked up without restarting. All standard directives are supported:
Host myserver
HostName 10.0.1.50
User deploy
Port 2222
IdentityFile ~/.ssh/id_ed25519
Then just connect with host: "myserver" — the alias, hostname, user, port, and key are all resolved automatically.
Tool Reference
ssh_connect
| Parameter | Type | Required | Description |
|---|---|---|---|
host |
string | yes | Hostname, IP, or SSH config alias |
username |
string | no | Override SSH config user |
port |
number | no | Override SSH config port (default 22) |
password |
string | no | Password auth |
privateKeyPath |
string | no | Path to private key (e.g. ~/.ssh/id_ed25519) |
passphrase |
string | no | Passphrase for encrypted private key |
useAgent |
boolean | no | Use SSH agent (default: true if no other auth given) |
connectTimeout |
number | no | Timeout in ms for TCP + handshake (default 20000) |
connectionId |
string | no | Custom ID for this connection (default: host value) |
Returns a connection ID used by the other tools.
ssh_execute
| Parameter | Type | Required | Description |
|---|---|---|---|
connectionId |
string | yes | Connection ID from ssh_connect |
command |
string | yes | Shell command to run |
cwd |
string | no | Working directory (wraps as cd <cwd> && <command>) |
timeout |
number | no | Command timeout in ms (default 30000) |
errOnNonZero |
boolean | no | Mark non-zero exit as MCP error (default false) |
Returns stdout, stderr, and exit code. Output is capped at 1MB per stream.
errOnNonZero defaults to false because many commands like grep (no match = exit 1) and diff (differences = exit 1) use non-zero exits for normal results.
ssh_disconnect
| Parameter | Type | Required | Description |
|---|---|---|---|
connectionId |
string | yes | Connection ID to close |
ssh_list_connections
No parameters. Returns all connections with host, port, user, and status (connected/dead).
Broken Pipe Handling
- Keepalive packets are sent every 10 seconds to detect dead connections early
- If a connection drops,
ssh_executereturns a clear error telling you to reconnect - An outer timeout wraps the entire TCP + SSH handshake so connections to firewalled/black-hole IPs don't hang forever
- All connections are cleaned up on
SIGTERM/SIGINT
Example Session
1. ssh_connect({ host: "myserver" })
→ Connected to 10.0.1.50:2222 as deploy. Connection ID: "myserver"
2. ssh_execute({ connectionId: "myserver", command: "uname -a" })
→ Linux prod-01 5.15.0 #1 SMP x86_64 GNU/Linux
→ Exit code: 0
3. ssh_execute({ connectionId: "myserver", command: "ls", cwd: "/var/log" })
→ syslog auth.log kern.log ...
→ Exit code: 0
4. ssh_list_connections()
→ - myserver: 10.0.1.50:2222 as deploy [connected]
5. ssh_disconnect({ connectionId: "myserver" })
→ Disconnected "myserver".
Security Notes
- Host key verification is not enforced (TOFU — trust on first use). The server accepts any host key.
- Passwords and passphrases are passed as tool parameters and are not persisted, but may appear in MCP client logs depending on your client.
- The server runs with the permissions of the user who started it.
License
MIT
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。