AI agents. Under your control.
Sentinel is a distributed platform for AI agents. A Control Plane on your NAS hands work to workers on your machines or in the cloud, and nothing risky happens without your approval.
Orchestration that keeps the work in Git
A web dashboard, AgentD for macOS, Linux and Windows, and a native iOS client on the same API.
AgentD workers
Outbound connections from your machines, auto-start, reconnect, upgrades with rollback and live CPU and RAM in the dashboard.
Multiple runtimes
Codex, Claude Code and Cursor Agent, picked automatically with fallback based on what each worker supports.
Role workflows
Architect → Developer → Reviewer with an optional QA step, parallel branches and a fix-up loop.
Policy and approvals
Risky operations wait for approval. Every task runs in an isolated Git worktree at read or develop level.
Scheduled tasks
Triggered by cron or Git events, with run history and pull requests on GitHub and GitLab.
Costs under control
Usage overview, monthly limits, and paid cloud runs only with explicit consent.
The Control Plane never gets more than the worker allows.
Effective permissions are the intersection of the task request, the agent policy and the local policy on the machine. Worker identity is bound to a key in the keychain, and every decision in the audit log has a name on it.
- Public access
- Cloudflare Tunnel
- Workers
- Tailscale · WireGuard
- Users
- admin · operator · reader
Runs on your hardware.
The Control Plane and workers run on your own hardware. Code stays in your repositories and only goes to the cloud when the repository data policy allows it.
Requirements
- Control Plane
- Docker (QNAP Container Station) or .NET 10
- AgentD
- macOS arm64, Linux x64, Windows x64
- Runtimes
- Codex, Claude Code or Cursor Agent
- iOS client
- iOS 17 or later, Face ID
- Language
- Czech