MCP SSH Proxy
Quinbook/McpSshProxy
Proxy SSH com aprovação humana para agentes IA — execute com segurança
Ver no GitHubPonte SSH controlada por humanos para agentes IA, com aprovação obrigatória antes de executar comandos. Oferece execução remota, transferências SFTP e terminais interativos multi-servidor, com credenciais criptografadas e detecção automática de comandos perigosos.
Instalação
npx convez-ai Quinbook/McpSshProxy1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Agentes IA executando tarefas em servidores de produção com controle humano
- DevOps automatizados com camada de segurança e auditoria
- Transferências SFTP seguras entre sistemas com revisão prévia
Veredito

Ferramenta sólida para quem precisa dar autonomia a agentes IA sem perder controle sobre infraestrutura crítica. A arquitetura de aprovação humana-em-loop é bem implementada e reduz risco de operações destrutivas acidentais.
README
A tradução para o português deste README ainda está na fila.
MCP SSH Proxy
A human-in-the-loop SSH bridge for AI agents. It exposes an MCP server with two tools — ssh_list_hosts and ssh_exec — but every command an agent wants to run is held for your review in a desktop app before it touches a server. You see the exact command and the target host, then click Run, edit it, or Reject it.
Servers are configured only in the app's UI — nothing is hardcoded, and credentials are encrypted at rest with the OS keystore (Windows DPAPI via Electron safeStorage).
Sibling project to MCP SQL Proxy — same approval-loop architecture, applied to SSH instead of SQL.
Features
- 🔐 Human-in-the-loop approval — every agent command is shown for review (Run / edit / Reject) before it runs; the result is returned to the agent automatically.
- ⚙️ Per-command confirm modes — a global confirm-each ⟷ auto-run switch, overridable per server (
always confirm/auto). - 🖥️ Built-in interactive terminals — full PuTTY-style xterm.js shells, multiple servers at once in tabs; type freely, see colors/vim/live output. Agent commands are echoed into the matching server's terminal as a labelled blue bar.
- 🗂️ Multi-server aware — pending requests are scoped to the selected server with a
!badge on the others; idle target servers are auto-connected on demand. - 📤 SFTP file transfer —
ssh_upload/ssh_downloadmove files to/from a server (inline content or a local path), each approved in the app. - 🔑 Any key format — OpenSSH, PEM/PKCS#8, and PuTTY
.ppk(v2 & v3) are detected and converted automatically; key, password, or ssh-agent auth. - 📥 PuTTY session import — pull saved sessions (host/port/user/key) straight from the Windows registry.
- 🕑 Persistent per-server history — every executed command is stored across sessions.
- 🛑 Danger mode — destructive commands (
rm,dd,shutdown,sudo,systemctl stop, redirects into/, …) flash the window red.
Why
Letting an agent run arbitrary shell commands on a production server is risky. This proxy keeps a person in the loop: the agent proposes, you approve. Destructive commands (rm, dd, shutdown, sudo, systemctl stop, redirects into /, …) light the whole window up in danger mode so a careless approval is hard.
Architecture
Claude / MCP client ──stdio──► MCP server ──WebSocket(127.0.0.1:52346)──► Electron app
(ssh_exec) (dist/mcp/server.js) (approval UI + ssh2)
- MCP server (
src/mcp/server.ts) — speaks MCP over stdio, forwards each request to the desktop app and waits for the result. Auto-launches the app if it isn't running. - Electron app (
src/electron/main.ts) — holds the server list, renders the approval queue, and runs approved commands viassh2. Single-instance: many MCP clients share one window. - Renderer (
src/renderer/index.html) — the UI: server list, per-server approval panel, tabbed interactive terminals (xterm.js), and all-server history.
Nothing binds to anything but 127.0.0.1.
Tools
| Tool | Approval? | Description |
|---|---|---|
ssh_list_hosts | no | Returns the configured host names (+ address/user/auth). No secrets. Call this first. |
ssh_exec | yes | Runs a command on a host (by name). You approve/edit/reject in the app. Returns stdout, stderr, exit code. |
ssh_upload | yes | Uploads a file via SFTP — inline content (text/base64) or a local_path on the proxy machine → remote_path. |
ssh_download | yes | Downloads a file via SFTP — returned inline (UTF-8 or base64, capped) or saved to a local_path. |
Install & build
npm install
npm run build
Run the desktop app standalone
npm start
Open Servers → + Add and configure a host:
- Name — the handle the agent uses (e.g.
web-01) - User / Host / Port
- Authentication — Private Key (with optional passphrase), Password, or SSH Agent
Hit Test Connection to verify. Secrets you type are encrypted before they are stored; the UI never reads them back in clear text (leave a secret field blank when editing to keep the stored value).
Register as an MCP server
Point your MCP client at the built server entry (dist/mcp/server.js). For Claude Code:
claude mcp add ssh-proxy -- node /absolute/path/to/McpSshProxy/dist/mcp/server.js
Or in a client config:
{
"mcpServers": {
"ssh-proxy": {
"command": "node",
"args": ["/absolute/path/to/McpSshProxy/dist/mcp/server.js"]
}
}
}
The first ssh_exec call auto-launches the approval window.
Security notes
- All traffic is loopback-only (
127.0.0.1:52346). - Passwords and key passphrases are encrypted with the OS keystore (
safeStorage). On platforms without an available keystore they fall back to obfuscated-but-not-encrypted local storage — prefer key/agent auth there. - The agent can only target servers you have configured by name. An
ssh_execfor an unknown host is refused. - This is an approval tool, not a sandbox: once you click Run, the command executes with the configured user's privileges. Review before approving.
License
MIT — see LICENSE.