Run cost
Your default wallet. free = only tools that cost nothing per use (your own subscriptions, local execution). paid_ok = paid frontier models and tools are allowed too (cost is disclosed each time). Model routing for planning follows this setting.
Scheduler
Fires automations from cron and build-state events.
Local execution (autorun)
Whether the hub runs local agents inside its own process. Turn it off and work only runs when a dedicated worker is up (broker-only).
Model routing
Where each step of a plan goes. cheap = light work (free subscription or local), strong = work that matters (your own Claude, etc). Blank = default.
Provider details (ollama / openai / anthropic)
Default host/model on the runner side (changes apply after a hub restart). API keys cannot be set here — env / .dev.vars only.
Automations
Saved flows running on their own (schedule / build-state). Turning one off stops it firing — it is not deleted.
Templates
Install a bundled flow in one click. It is copied in as an editable workflow.
Outcome checks
Judges whether an automation produced what you expected, and alerts you when it did not. assert = deterministic ($0), judge = an AI answers yes/no.
Drift detection
A louder warning than an outcome check: recorded when the same automation fails three times running, or its output shape changes suddenly (detected automatically, read-only).
Suggestions from Shenron
Shenron watches its own run history and offers things like "this could run on a schedule" or "this keeps failing". It does not read any external inbox.
Environment & API keys
Shows only whether a key is present — never the value. The LLM can also run locally through claude -p at no per-use cost.