managed cloud · open-source agent · Apache 2.0

Connect AI to
your servers.

Name: SentinelX
Author: pensa.ar

SentinelX connects Claude, ChatGPT and any MCP client to your Linux, macOS and Windows servers. One command per host, Google sign-in, your whole fleet in one dashboard.

↓ Start free ↗ See dashboard ↗ GitHub

$ curl -fsSL https://get.sentinelx.app | bash

One command — auto-detects Linux or macOS.

PS> iwr -useb https://get.sentinelx.app/install.ps1 -OutFile "$env:TEMP\sx.ps1"
    powershell -ExecutionPolicy Bypass -File "$env:TEMP\sx.ps1"

PowerShell — needs Python 3.12+ and git. Add -User for a no-admin install.

Then enroll: mcp.sentinelx.app → "+ Enroll a new host" → paste token. Done.

// available on GitHub MCP Registry ChatGPT

works with Claude VS Code Cursor Cline Any MCP client

// how it works

An agent on each host. A hub that handles the rest.

The hub handles OAuth, identity and MCP. The agent only connects outbound — no inbound ports, no DNS, no certificates on your host.

Claude · ChatGPT · Cursor

any MCP client

↓HTTPS + OAuth · Google sign-in

mcp.sentinelx.app

managed hub — OAuth, MCP, dashboard

↓WebSocket · outbound from the agent

sentinelx-cloud-core

open-source agent on your server

Your servers never expose an inbound port to the internet.

01 — INSTALL

One command, one minute

Run the installer. It creates a system user, registers a systemd service, and connects out to the hub.

02 — ENROLL

Sign in with Google

03 — DASHBOARD

See your fleet

Every host with status, version and last-seen. Label, kick sessions, or unenroll — from the browser at mcp.sentinelx.app/dashboard.

04 — USE IT

From any AI client

Add the hub as an MCP connector in Claude or ChatGPT. The AI targets hosts by name — and only runs what the allowlist permits.

// AI clients

Works with any MCP client.

One hub URL, any MCP client. Sign in once with Google — every client you connect sees the same fleet.

Claude (claude.ai, Claude Desktop, Claude Code)

ChatGPT (in the app directory ↗)

VS Code

Cursor

Cline

Any MCP-compatible agent or script

// connect Claude in 30 seconds

1. claude.ai → Settings → Connectors → Add custom connector
2. URL: https://mcp.sentinelx.app/mcp/mcp
3. Sign in with Google when prompted
4. Ask Claude: "list my hosts"

// connect ChatGPT in 30 seconds

1. Open SentinelX in the ChatGPT app directory ↗
2. Click Connect and sign in with Google
3. Ask ChatGPT: "list my hosts"

// tools the AI sees, per host

sentinel_state host facts

sentinel_capabilities what's allowed

sentinel_exec run an allowlisted command

sentinel_edit structured file edits

sentinel_script_run temp bash / python script

sentinel_service start / stop / restart / reload

sentinel_upload_file upload to host

sentinel_list_hosts multi-host fleet view

…and more — see the agent's docs

// open standards, no vendor lock-in

protocolModel Context Protocol (MCP)

transportStreamable HTTP

authOAuth 2.0 / OIDC + DCR

identityGoogle sign-in

agentApache 2.0-licensed, open source

// beyond ops

Persistent dev sessions, with memory.

Most AI coding tools start every session from zero. SentinelX runs inside your chat, so context and decisions accumulate across days. Add git on the host and the AI actually remembers your project.

01 — REFACTOR OVER DAYS

Iterate on a repo across sessions

Clone once, refactor over a week. Thursday's session remembers what Monday tried.

02 — DEBUG WITH HISTORY

Bugs that need yesterday's context

"Remember the timeout from last week? It came back." The chat holds the thread; the agent re-runs the diagnostics live.

03 — DECISIONS THAT STICK

Architecture rationale stays decided

Why Postgres over SQLite, why the worker is on its own host. The reasoning lives in the conversation, not in comments you'll forget to grep.

// dogfooded

This site was built and is continuously updated with SentinelX.

The agent edits /var/www/sentinelx.pensa.ar/index.html directly, with git tracking every change. Including this paragraph.

// security model

Least privilege, by design.

The agent is open source — audit exactly what runs on your server.

🔒

Command allowlist

Only commands explicitly listed in the policy run. Paths and arguments are validated too.

👤

Isolated service user

A dedicated sentinelx user, no login shell, sudo scoped to exactly what the policy permits.

🎟

OAuth via Google

Google sign-in, JWT sessions with explicit scopes. The agent only trusts the hub.

🔑

Outbound only

One outbound WebSocket. No inbound ports, no public DNS, no certificates. Your firewall stays closed.

📋

Audit log

Every operation logged: user, host, command, output, return code. Blocked attempts too.

🧰

Open-source agent

The agent is Apache 2.0-licensed and lives at pensados/sentinelx-cloud-core. Read it, fork it, run a fork — it's yours.

// editor & CLI clients

Drop-in config for JSON-based clients.

Same hub URL, one paste. The only thing that changes per client is the top-level key and how the transport field is spelled — get that wrong and the client silently falls back to the legacy SSE transport.

VS Code · ~/.config/Code/User/mcp.json

{
  "servers": {
    "sentinelx": {
      "type": "http",
      "url": "https://mcp.sentinelx.app/mcp/mcp"
    }
  }
}

Top-level key servers, transport "http".

Cline · cline_mcp_settings.json

{
  "mcpServers": {
    "sentinelx": {
      "type": "streamableHttp",
      "url": "https://mcp.sentinelx.app/mcp/mcp"
    }
  }
}

Transport must be "streamableHttp" — camelCase, no hyphen. Omitting it or using "http" throws Invalid schema and falls back to SSE (405).

Cursor · ~/.cursor/mcp.json

{
  "mcpServers": {
    "sentinelx": {
      "url": "https://mcp.sentinelx.app/mcp/mcp"
    }
  }
}

url only — Cursor auto-detects transport and handles OAuth. Heads-up: Cursor caps ~40 active tools and SentinelX exposes 38, so disable tools you don't use or trim other servers.

Any MCP client · the common shape

{
  "mcpServers": {
    "sentinelx": {
      "url": "https://mcp.sentinelx.app/mcp/mcp"
    }
  }
}

Most clients use mcpServers + url. A few differ — check your client's docs for the transport key (e.g. type or serverUrl).

Android Studio — known issue. Its built-in MCP client has a PKCE bug (the code_verifier doesn't match its own code_challenge), so the OAuth handshake fails on the client side. Nothing to fix server-side — use VS Code or another client until JetBrains ships a fix.

Windsurf — known issue. Windsurf's MCP client validates the OAuth issuer in the wrong place (it expects it in the token response body instead of the RFC 9207 callback parameter it already receives), so the token exchange fails on the client side. The config is correct — it'll connect once Windsurf ships the upstream fix (rust-sdk#896). Use Cursor or VS Code in the meantime.

// one URL for every client
https://mcp.sentinelx.app/mcp/mcp — sign in with Google on first connect.

// self-hosted alternative

Want to run your own hub?

Prefer running the whole stack yourself, OAuth included? The Docker edition bundles agent, MCP bridge and Keycloak. Same protocol, same security model.

SentinelX Docker self-hosted

One-line install, 4 containers on your own host. You bring DNS, certificates and ops. For full data sovereignty or air-gapped setups.

$ curl -fsSL https://raw.githubusercontent.com/pensados/sentinelx-docker/main/install.sh | bash

↗ sentinelx-docker ↗ Setup guide

Connect AI toyour servers.