When you submit a message to the optional AI assistant, Varc sends that message and a bounded snapshot of the current Console command, script, dashboard, or widget to our API. Cloudflare Workers AI processes that material to generate the reply. Nothing is sent merely because an AI button is visible or the assistant is open.
The snapshot is the text you are working on, sent as you wrote it: the Console command in the editor, the script source, or the scripts and titles of the widgets on the dashboard. If you have typed a key name, a value, or a password into that text — SET session:abc …,CONFIG SET requirepass … — it goes with it. Alongside it we send the deployment type, the logical database number, the names of the commands your server supports, local parse diagnostics, and the public scripting API description.
What is never sent is everything Varc read on your behalf: the connection details and credentials in your OS keychain, environment-variable values, command results, script output, transformer samples, and widget results or chart history. The assistant sees what you typed, never what your server answered.
Varc does not write prompts, source context, or replies to D1, logs, analytics, exception payloads, or traces. We retain only content-free usage records such as a hashed licence or Team subject, UTC month, request count, surface, status, latency, model id, and aggregate token usage. Conversations remain in app memory and are cleared when you close the assistant or workspace, quit Varc, or disable AI features.
You can turn the feature off under Settings → AI assistant. This removes every AI trigger, cancels current client waits, closes the assistant, and clears its in-memory conversations.