SSH MCP Server
Enables AI to execute commands on remote hosts via SSH, supporting password and key authentication.
README
SSH MCP Server
A minimal MCP server that exposes one tool — ssh_exec — so Claude Code, OpenAI Codex, or any MCP-compatible AI can run commands on a remote host over SSH and inspect the output. It supports both stdio and streamable-http transports.
What it does
- Starts an MCP stdio server by default.
- Can also serve MCP over
streamable-httpat/mcp. - Exposes
ssh_exec(command, host, user, password, port, identity_file, timeout_sec, strict_host_key_checking). - Uses paramiko for SSH — no dependency on a system
sshbinary. - Supports password authentication and public key authentication.
- Returns structured output:
ok,exit_code,stdout,stderr,target.
Prerequisites
- Docker
- Claude Code CLI (
claude) and/or OpenAI Codex CLI (codex) installed and authenticated
Build the image
cd /path/to/ssh-mcp
docker build -t ssh-mcp:latest .
Run as Streamable HTTP
Set MCP_TRANSPORT=streamable-http to start an HTTP MCP endpoint instead of the default stdio server.
Local Python
python3 -m venv .venv
. .venv/bin/activate
pip install -r requirements.txt
MCP_TRANSPORT=streamable-http \
MCP_HOST=127.0.0.1 \
MCP_PORT=8000 \
python server.py
The MCP endpoint will be available at http://127.0.0.1:8000/mcp.
Docker
When running in Docker, bind to 0.0.0.0 so the port is reachable outside the container:
docker run --rm -p 8000:8000 \
-e MCP_TRANSPORT=streamable-http \
-e MCP_HOST=0.0.0.0 \
-e MCP_PORT=8000 \
-e SSH_DEFAULT_HOST=your-host.example.com \
-e SSH_DEFAULT_USER=ubuntu \
-e SSH_ALLOWED_HOSTS=your-host.example.com \
ssh-mcp:latest
The MCP endpoint will be available at http://localhost:8000/mcp.
If your client supports stateless HTTP MCP sessions, add -e MCP_STATELESS_HTTP=true to the docker run command.
Add to Claude Code
Option 1: Claude Code CLI (recommended)
These examples use the default stdio transport.
Key-based auth:
claude mcp add sshRemote -- \
docker run --rm -i \
-v "$HOME/.ssh:/home/app/.ssh:ro" \
-e HOME=/home/app \
-e SSH_DEFAULT_HOST=your-host.example.com \
-e SSH_DEFAULT_USER=ubuntu \
-e SSH_ALLOWED_HOSTS=your-host.example.com \
ssh-mcp:latest
Password auth:
claude mcp add sshRemote -- \
docker run --rm -i \
-e SSH_DEFAULT_HOST=your-host.example.com \
-e SSH_DEFAULT_USER=ubuntu \
-e SSH_DEFAULT_PASSWORD=your-password \
-e SSH_ALLOWED_HOSTS=your-host.example.com \
ssh-mcp:latest
Verify it was added:
claude mcp list
Option 2: ~/.claude.json (manual config)
Add the following to your ~/.claude.json under the mcpServers key:
{
"mcpServers": {
"sshRemote": {
"command": "docker",
"args": [
"run", "--rm", "-i",
"-v", "/Users/<your-user>/.ssh:/home/app/.ssh:ro",
"-e", "HOME=/home/app",
"-e", "SSH_DEFAULT_HOST=your-host.example.com",
"-e", "SSH_DEFAULT_USER=ubuntu",
"-e", "SSH_ALLOWED_HOSTS=your-host.example.com",
"ssh-mcp:latest"
]
}
}
}
Option 3: Project-level config (.claude/settings.json)
To scope the server to a single project, add it to .claude/settings.json in your project root:
{
"mcpServers": {
"sshRemote": {
"command": "docker",
"args": [
"run", "--rm", "-i",
"-v", "/Users/<your-user>/.ssh:/home/app/.ssh:ro",
"-e", "HOME=/home/app",
"-e", "SSH_DEFAULT_HOST=your-host.example.com",
"-e", "SSH_DEFAULT_USER=ubuntu",
"-e", "SSH_ALLOWED_HOSTS=your-host.example.com",
"ssh-mcp:latest"
]
}
}
}
Add to OpenAI Codex
Option 1: Codex CLI (recommended)
These examples use the default stdio transport.
codex mcp add sshRemote -- \
docker run --rm -i \
-v $HOME/.ssh:/home/app/.ssh:ro \
-e HOME=/home/app \
-e SSH_DEFAULT_HOST=your-host.example.com \
-e SSH_DEFAULT_USER=ubuntu \
-e SSH_ALLOWED_HOSTS=your-host.example.com \
ssh-mcp:latest
codex mcp list
Option 2: ~/.codex/config.toml
[mcp_servers.sshRemote]
command = "docker"
args = [
"run", "--rm", "-i",
"-v", "/Users/<your-user>/.ssh:/home/app/.ssh:ro",
"-e", "HOME=/home/app",
"-e", "SSH_DEFAULT_HOST=your-host.example.com",
"-e", "SSH_DEFAULT_USER=ubuntu",
"-e", "SSH_ALLOWED_HOSTS=your-host.example.com",
"ssh-mcp:latest"
]
startup_timeout_sec = 20
tool_timeout_sec = 120
Then restart Codex (CLI/TUI/IDE extension) and verify:
codex mcp list
Specifying host, user, and port
You have two ways to provide connection details:
Via environment variables (defaults for every call)
Pass -e flags to docker run:
| Variable | Example | Description |
|---|---|---|
SSH_DEFAULT_HOST |
my-server.example.com |
Remote hostname or IP |
SSH_DEFAULT_USER |
ubuntu |
SSH login username |
SSH_DEFAULT_PORT |
2222 |
SSH port (default: 22) |
SSH_ALLOWED_HOSTS |
my-server.example.com |
Comma-separated allowlist of hosts |
Via tool arguments (per-call override)
You can also pass connection details directly when prompting the AI:
Run ssh_exec with host="10.0.1.50", user="admin", port=2222, command="uptime"
Per-call arguments override the environment variable defaults.
Authentication
The ssh_exec tool selects the auth method based on what you provide:
| What you provide | Auth method used |
|---|---|
password |
Password auth only (key lookup disabled) |
identity_file |
The specified private key file |
| Neither | Keys from ~/.ssh and the SSH agent, tried automatically |
Password authentication
Pass the password as an environment variable default or as a per-call argument:
As a default (env var):
-e SSH_DEFAULT_PASSWORD=hunter2
As a per-call argument:
Run ssh_exec with host="10.0.1.50", user="admin", password="hunter2", command="uptime"
When a password is supplied, key lookup and the SSH agent are disabled so the auth method is unambiguous.
Public key authentication
Mount your ~/.ssh directory into the container and the server will automatically try the keys it finds there (including the SSH agent if a socket is forwarded):
-v "$HOME/.ssh:/home/app/.ssh:ro"
-e HOME=/home/app
To use a specific key file instead of scanning ~/.ssh, pass it as a per-call argument:
Run ssh_exec with host="10.0.1.50", user="admin", identity_file="/home/app/.ssh/id_deploy", command="uptime"
Setup (if you don't already have a key pair):
-
Generate a key:
ssh-keygen -t ed25519 -C "ssh-mcp" -
Copy it to the remote host:
ssh-copy-id -i ~/.ssh/id_ed25519.pub user@your-host.example.com -
Test passwordless login:
ssh user@your-host.example.com "echo ok"
Environment variables reference
| Variable | Default | Description |
|---|---|---|
SSH_DEFAULT_HOST |
(none) | Default remote host |
SSH_DEFAULT_USER |
(none) | Default SSH username |
SSH_DEFAULT_PASSWORD |
(none) | Default SSH password (password auth) |
SSH_DEFAULT_PORT |
22 |
Default SSH port |
SSH_DEFAULT_TIMEOUT_SEC |
60 |
Default command timeout (seconds) |
SSH_CONNECT_TIMEOUT_SEC |
10 |
SSH connection timeout (seconds) |
SSH_ALLOWED_HOSTS |
(none, allow all) | Comma-separated host allowlist |
SSH_MAX_OUTPUT_CHARS |
12000 |
Output truncation limit |
MCP_TRANSPORT |
stdio |
MCP transport: stdio or streamable-http |
MCP_HOST |
127.0.0.1 |
Bind host for HTTP transports |
MCP_PORT |
8000 |
Bind port for HTTP transports |
MCP_STREAMABLE_HTTP_PATH |
/mcp |
HTTP path for the MCP endpoint |
MCP_STATELESS_HTTP |
false |
Enable stateless streamable HTTP sessions |
MCP_MOUNT_PATH |
/ |
Advanced: ASGI mount path used by FastMCP |
Example prompts
- "Use
sshRemotessh_execand rununame -a." - "Run
df -hon the remote host and summarize disk usage." - "Run
journalctl -u nginx -n 200 --no-pagerand find recent errors." - "Check CPU load on
my-server.example.comwithuptime."
Security notes
- This server can execute arbitrary remote shell commands — treat it as a high-privilege tool.
- Always set
SSH_ALLOWED_HOSTSto restrict which hosts can be targeted. - Use a least-privileged SSH user on the remote host where possible.
- Keep
strict_host_key_checking=true(the default) to prevent MITM attacks. - Prefer key-based auth over passwords where possible. If you do use a password, pass it via an environment variable rather than as a tool argument so it stays out of conversation logs.
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。