How self-driving works
What Aloof does on its own today, with the exact limits it uses. These numbers come straight from the platform code.
Bad deploys roll themselves back
For 5 minutes after every deploy, Aloof compares the new release with the previous one. If the 5xx rate is more than 5 points higher and more than double, or p95 latency is more than double and over 500 ms slower, it rolls back with zero downtime and records why. You see the reason in release history, the CLI, the dashboard, over MCP and in the audit log. You can turn it off per app (set_self_heal over MCP).
Ask your assistant why it broke
The diagnose_app MCP tool gathers recent errors, failed health checks, deploys, rollbacks, restarts and exit codes, then names the likely cause and fix. Your assistant explains it in plain language. It's read-only and redacts secrets. The causes come from rules, not an AI model.
Traffic routes around failing instances
Requests go to the instance with the fewest open connections. Five failures in a row (connection errors, timeouts, 502 to 504) take an instance out for 10 seconds, backing off up to 5 minutes. Two failed health checks also take it out. Aloof never removes more than half your instances, or the last one. This applies when you run 2 or more instances.
Deploys that only switch over when healthy
Redeploys only take traffic after the new instance passes its health check, with zero downtime. Health checks run every 5 seconds and replace unhealthy instances automatically. Roll back to any of your last 20 releases with one command.
Scales itself, down to zero
Autoscaling checks in-flight requests every 3 seconds. Idle apps sleep and wake on the next request in about a second, and you pay nothing while they sleep.
Your AI assistant has the wheel
39 MCP tools let Claude, ChatGPT or Cursor deploy, scale, debug and launch templates. A remote MCP server with OAuth 2.1 gives them scoped tokens you can revoke at any time. There's also llms.txt and an OpenAPI spec.
One-click templates
Deploy n8n, uptime-kuma, umami, vaultwarden, ntfy, pocketbase, linkding in one click from templates, or add a “Deploy on Aloof” button to your own repo's README.
Security defaults
What the platform does today. We hold no certifications yet.
- TLS everywhere. Every app, API and MCP endpoint is served over HTTPS with automatically issued certificates. Certificates are only requested for hosts we know.
- Encrypted secrets. Environment variables are encrypted at rest and never shown back. API keys are stored only as SHA-256 hashes.
- Isolated containers. Each app runs in its own container with CPU and memory limits. Nodes join over WireGuard and expose no inbound ports.
More on the security page. Found an issue? Email security@aloof.si.
Coming soon
Not live yet: predictive scaling, auto right-sizing, AI auto-fix, GPUs (waitlist) and more regions.
FAQ
What does “self-driving” mean today?
Aloof rolls back deploys that raise errors or latency, replaces instances that fail health checks, routes traffic away from failing instances and autoscales, all without you stepping in. Predictive scaling, auto right-sizing and AI auto-fix are coming soon and not live yet.
Will automatic rollback cover my app?
Only if it gets enough traffic in the 5 minutes after a deploy: at least 100 requests, with at least 10 bad requests from at least 3 IPs. Low-traffic apps aren't covered, but health-gated deploys and one-command rollback still are. You can turn automatic rollback off per app.
Does diagnose_app send my data to an AI model?
No. It's a read-only MCP tool that applies rules to your app's recent errors, health checks, deploys, rollbacks, restarts and exit codes, and redacts secrets. Your own assistant reads the result and explains it.
Can an AI agent deploy for me safely?
Yes. Agents use OAuth 2.1 with PKCE and get scoped tokens that you can revoke at any time.
What can I deploy?
Any Docker image, or a git repo with a Dockerfile. This covers Node.js, Python, Go, Rust, Ruby, Java/JVM, PHP and static sites.
How does scale-to-zero affect latency?
Idle apps sleep and wake on the next request in about a second.
How are secrets handled?
Environment variables are encrypted at rest and never shown back. API keys are stored only as SHA-256 hashes.
Where does my code run?
On Hetzner Cloud in Nuremberg, Germany. The control plane runs there too, and encrypted off-site backups are stored in Helsinki, Finland. More regions are coming soon.
Do you support GPUs?
Not yet. GPUs are coming soon: L4, RTX 4090, L40, A100 80GB and H100, billed per second with scale to zero. Join the waitlist.