mcp-ssh-agentic
An MCP server for SSH/SCP operations with passwordless authentication, enabling remote command execution, file transfer, and session management.
README
mcp-ssh-agentic
An MCP server for agentic SSH/SCP operations with passwordless authentication. It uses the operating system's native ssh and scp binaries, so existing SSH features continue to work seamlessly, including public key authentication, ~/.ssh/config, ssh-agent, ProxyJump, and known_hosts.
Connection Efficiency
On Linux and macOS, each target uses SSH ControlMaster multiplexing:
- Socket:
~/.cache/mcp-ssh-agentic/mux/<hash> ControlPersist=600(master connection stays alive for 10 minutes of idle time)- Subsequent calls to the same host reuse the existing TCP connection and authentication
- Stale sockets are automatically removed and the connection is retried once
ssh_closeexplicitly closes the master connection
Windows (cmd / PowerShell / Git Bash)
Native Win32-OpenSSH does not support ControlMaster (this affects all three common Windows shells — they share the same ssh.exe). Multiplexing is disabled by default on win32; every tool call opens a fresh SSH connection. Behavior is otherwise the same.
| Client | Mux default | Notes |
|---|---|---|
| Windows cmd.exe | off | Uses Win32-OpenSSH |
| Windows PowerShell / pwsh | off | Same OpenSSH client |
| Git Bash | off | Usually the same Win32-OpenSSH on PATH |
| WSL (Ubuntu, etc.) | on | Runs as Linux — preferred on Windows for connection reuse |
| macOS / Linux | on | Full ControlMaster support |
If a ControlMaster error is still seen (for example after forcing mux on), the server auto-disables multiplexing for the rest of the process and retries without it.
Override with env:
{
"mcpServers": {
"ssh-agentic": {
"command": "npx",
"args": ["-y", "@jahrulnr/mcp-ssh-agentic"],
"env": {
"MCP_SSH_AGENTIC_MUX": "0"
}
}
}
}
MCP_SSH_AGENTIC_MUX=0— force off (also:false,no,off)MCP_SSH_AGENTIC_MUX=1— force on (only if yoursshactually supports ControlMaster)
Running with npx
After a release, the package is on npmjs and GitHub Packages as @jahrulnr/mcp-ssh-agentic.
npmjs (simplest):
{
"mcpServers": {
"ssh-agentic": {
"command": "npx",
"args": ["-y", "@jahrulnr/mcp-ssh-agentic"]
}
}
}
GitHub Packages (needs a PAT with read:packages in ~/.npmrc):
@jahrulnr:registry=https://npm.pkg.github.com
//npm.pkg.github.com/:_authToken=YOUR_GITHUB_PAT
{
"mcpServers": {
"ssh-agentic": {
"command": "npx",
"args": ["-y", "--registry=https://npm.pkg.github.com", "@jahrulnr/mcp-ssh-agentic"]
}
}
}
Targets must always be specified as user@host[:port], for example demo@127.0.0.1:22. The port may also be configured through aliases in ~/.ssh/config.
Available Tools
ssh_ping, ssh_read_file, ssh_write_file, ssh_read_image, ssh_list_dir, ssh_mkdir, ssh_grep, ssh_apply_patch, ssh_delete, ssh_exec, ssh_exec_result, ssh_exec_kill, ssh_interactive_exec, ssh_interactive_input, ssh_interactive_close, ssh_interactive_list, ssh_scp_to, ssh_scp_from, ssh_close
Examples:
ssh_read_file("demo@127.0.0.1:22", "/etc/hostname")
ssh_write_file("demo@server", "/srv/app/.env", "PORT=3000\n")
ssh_mkdir("demo@server", "/srv/app/releases/42")
ssh_list_dir("demo@server", "/var/log")
ssh_grep("demo@server", "TODO", "/srv/app", "*.js")
ssh_exec("demo@server", "systemctl --user status my-service")
# Detached / background command:
ssh_exec("demo@server", "./build.sh", background=true)
# -> job_id=abc123
ssh_exec_result(job_id="abc123", wait=true, timeout_ms=60000)
ssh_exec_kill(job_id="abc123", signal="SIGTERM", cleanup=true)
# Commands that may require interactive input (sudo password, y/N prompts, setup wizards):
ssh_interactive_exec("demo@server", "sudo apt-get upgrade")
# -> session_id=abc123, status=running, output contains "[sudo] password for demo:"
ssh_interactive_input(session_id="abc123", input="secret123")
# -> additional output, e.g. "Do you want to continue? [Y/n]"
ssh_interactive_input(session_id="abc123", input="Y")
# -> status=exited (code 0) once the process finishes
ssh_interactive_close("abc123") # close manually if needed
ssh_scp_to("demo@server", "./dist/app.tar.gz", "/apps/app.tar.gz")
ssh_scp_from("demo@server", "/apps/backups/db.sql.gz", "./db.sql.gz")
ssh_close("demo@server")
Behavior Notes
ssh_deleteusesrm -ffor files andrm -rfonly whenrecursive=true.ssh_write_filewrites or overwrites a remote file directly from text (without creating a temporary local file). Useappend=trueto append instead of overwrite. Parent directories are created automatically unlesscreate_dirs=false.ssh_mkdiris equivalent tomkdir -pon the remote host, which is especially useful beforessh_scp_to, sincescpdoes not automatically create remote parent directories.ssh_exechas a default timeout of 30 seconds and a maximum output size of 5 MiB.ssh_read_imagesupports files up to 20 MiB.ssh_write_fileaccepts content up to 5 MiB. SCP operations default to a 120-second timeout.ssh_execalso acceptscwd,env,ok_codes, andmax_output_bytes.- Set
background=trueonssh_execto run a command detached and get ajob_id. Usessh_exec_resultto poll or wait, andssh_exec_killto stop or clean up the job log directory. - All remote commands (not just
ssh_exec) run through a non-login, non-interactive shell (bash --noprofile --norc -c, falling back tosh -c). This avoids failures caused by broken/etc/profile.dscripts on some servers and ensures consistent behavior across all tools.ssh_execalways returnsexit_code=Nalong with stdout. If stderr is present, it is included in a[stderr]section. Non-zero exit codes setisError, but stdout is still returned. - Interactive sessions (TTY/PTY):
ssh_interactive_execforces PTY allocation on the remote side (ssh -tt), allowing programs that require a real terminal (sudo,passwd, confirmation prompts, setup wizards, REPLs, etc.) to behave correctly even though the local side uses ordinary pipes. The server waits until output has been quiet forquiet_ms(default: 500 ms) or the process exits, then returns the collected output along with asession_id. Continue the session usingssh_interactive_input(leaveinputempty to simply wait for more output without sending anything). This mechanism is based on output inactivity rather than prompt detection. Commands that continuously produce output (such as long build logs) may cause the tool call to wait longer depending onquiet_msandmaxWaitMs. The server allows up to 8 concurrent interactive sessions, automatically cleans up sessions after 10 minutes of inactivity, and terminates all active sessions when the server exits. Usessh_interactive_listto view active sessions andssh_interactive_closeto close them manually. ssh_greptreats "no matches" as a successful result and still returns partial matches even if some paths cannot be read.ssh_scp_toandssh_scp_fromsupportrecursive=truefor directories. Local parent directories are created automatically when downloading. Remote parent directories must already exist before uploading (usessh_mkdirif needed).- The SSH and SCP clients are invoked with
-qto suppress MOTD and other unnecessary output on stderr.
Local Development
npm install
npm run check
npm test
npm start
Unit tests use createMockTransport() — the same SSH contract (exec / scp / close / spawnInteractive) executed in a local sandbox, without a real SSH host.
To test the MCP protocol, use MCP Inspector or any MCP client that supports stdio transport.
CI / Release
GitHub Actions (.github/workflows/ci.yml):
- Unit test (any branch/PR) — Node 18 / 22 / 24 →
npm run check+npm run test:unit - MCP test (after unit) — same Node matrix × (
nodebin |npxfromnpm pack) withMCP_SSH_AGENTIC_MOCK=1 - Push to
master— after both pass, if tagvX.Y.Zis new: create tag → publish to GitHub Packages + npmjs
Local: npm run test:all
Bump version in package.json before merging to master for a new release. Re-merging the same version skips tag/publish.
Secrets: NPM_TOKEN. GitHub Packages uses GITHUB_TOKEN (packages: write).
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。