Playground

The playground splits the screen in two: the sidebar on the left holds every knob that affects how the agent answers; the right pane renders the live widget exactly as your users will see it on your site.

Sidebar changes hit the preview the moment you save — no rebuild, no training run.

Training status pill

The pill at the top of the sidebar shows whether your agent is trained, currently training, or hasn’t been trained yet. The subtitle adds a timestamp and the size of the indexed corpus.

Compare AI models

Quick link into the side-by-side comparison view. Pick two models and run the same prompt against both — see Models Comparison.

Model

Tap the model row to open the picker — a searchable list grouped by provider (OpenAI, Anthropic, Google, xAI, Meta, DeepSeek, Moonshot). Locked models show greyed out with a tooltip explaining the upgrade required; available models highlight in mint.

Voice Mode

Pro-only toggle. When enabled, the widget exposes a microphone button and uses speech-to-text + text-to-speech for full voice conversations. See Settings → Voice for the model, transcriber, and voice-tuning options.

AI Actions

Empty-state link into the Actions surface. Once you’ve hooked up actions (lead capture, Calendly booking, Stripe checkout, etc.), they’ll surface here — and the agent can invoke them mid-conversation.

Instructions (System prompt)

The big textarea at the bottom. Pick a preset (Base Instructions, General AI agent, Customer support agent, Sales agent) and the prompt populates. Edit a single character and the dropdown switches to Custom — your changes are preserved.

The reset button next to the dropdown drops you back to the original preset. Useful when you’ve tweaked yourself into a corner.

The right pane renders your agent inside the same widget shell users will see — your colours, avatar, suggested queries, post-response follow-ups. Type into the input and the agent responds with the model and prompt currently in the sidebar.

During a training run the preview shows a spinner overlay with Training in progress— you can keep editing the configuration; it just won’t accept new chats until training finishes.

• Resize the browser — the preview keeps the widget proportions, useful for sanity-checking responsive behaviour
• Suggested queries from the chat-interface settings show up here too
• The widget always uses the saved widget config — preview doesn't reflect unsaved chrome changes

Edits in the sidebar go into a dirty state. A footer slides up from the bottom of the sidebar with two buttons:

• Discard — reverts every field back to the last saved state.
• Save to agent— writes the changes and updates the live widget. Saving doesn’t trigger a retrain unless you also changed sources.

The Compare button on the Compare AI Models card opens the side-by-side view — two (or more) playground panels rendered next to each other. Each panel has its own model picker; the shared Sync toggle fans your prompt out to every panel at once so you can see how each model responds to the same input.

Detail in Models Comparison.