Experimental features

Some features have unit tests but no end-to-end or real-infrastructure evidence (see feature status). They are off by default. Turn them on with one env var:

APP_EXPERIMENTAL=ai-models,load-balancer

The value is a comma separated list of feature keys. Unknown keys are logged at startup and ignored.

Key	Feature
ai-chat	In-app AI assistant chat and its settings
ai-models	AI models on GPU nodes
load-balancer	Load balancer across an app's replicas
iac	Platform as code: apply, diff, export
cloudflare-tunnel	Cloudflare Tunnel

What "off" means

• API: routes of the feature answer 404 with {"error": "...", "code": "experimental_feature_disabled", "feature": "<key>"}. The message names APP_EXPERIMENTAL.
• CLI: the commands are hidden from --help and from shell completion, and exit with a usage error naming APP_EXPERIMENTAL. The CLI reads its own environment, not the control plane's: set the variable in the shell you run the CLI from, and regenerate any installed completion script after changing it.
• MCP: the tools are not registered. The MCP server is a separate process and reads its own environment: set the variable there, not only on the control plane.
• Dashboard: navigation entries, command palette entries, g shortcuts and in-page pointers are hidden, pages redirect home on direct visits. The dashboard follows the control plane's setting.
• Reconcilers: the model, load balancer routing and Cloudflare Tunnel controllers do not start.
• Data: nothing is deleted. Turning a feature back on restores it as it was.

Set the variable on the control plane with sudo systemctl edit levelrail ([Service], then Environment=APP_EXPERIMENTAL=...) and restart.