Model
Pick which LLM answers — from a searchable list showing each model’s context window. Empty = the workspace default.
One block. Every option actually works.
Drop live AI chat onto any page, then choose the model, the tools, the agents, the layout, and the starter actions — and watch every setting take effect. Then run two versions against real traffic and keep the winner.
See it — pick a layout and a view
Layout — where the chat lives
View — how the empty state arranges
This layout is too narrow for a centered hero — it always uses Classic.
How it works
From one config object to the live conversation — nothing dropped on the way.
You configure the block
Every option is one field on a single config object (ChatWidgetConfig) — model, tools, agents, view, starter cards, features.
It reaches the live chat
On a page the block either renders the chat inline, or stamps its config so the always-mounted chat picks it up — no field silently dropped.
The server cross-checks it
Tool/agent allow-lists are re-checked against your stored workspace config — a forged request can only narrow, never widen, a restriction.
The runtime applies it
The picked model answers, excluded tools are filtered out, and the chat can only delegate to the agents you allowed.
What you can set
Six controls, one inspector. Empty always means “use the default”.
Tools
Restrict the chat to a set of tools. Empty means all enabled; a non-empty list actually filters the turn.
Agents
Choose which of your workspace’s agents (from slug/agents) the chat may delegate to.
Starter cards
Add quick-action chips. A card sends its prompt — or launches a workflow you attach to it.
Features
Toggle thread history, voice, speech, thinking, and starter suggestions per widget.
Header & placeholder
Set the title, logo, and input placeholder to match the surface it lives on.
How to use it
Six steps, no deploy.
- 1
Drop the block on a page
In the page editor, drag the AI Chat block onto any page — or edit your workspace’s default at /u/[slug]/studio/chat.
- 2
Pick a model (optional)
Search the model list and select one, or leave it blank to inherit the workspace default.
- 3
Scope tools & agents
Restrict what the chat can do and who it can hand off to. Empty = everything the workspace already allows.
- 4
Choose the look
Set the layout (icon, rail, panel, full, embedded) and the view (centered or classic).
- 5
Add starter cards
Give visitors one-click prompts — and attach a workflow to a card to run a real flow (book a demo, get a quote).
- 6
Publish
The live chat reflects every choice on the next turn. Nothing to redeploy.
Why it’s special
Configurable chat is common. Chat where every switch is real, safe, and testable is not.
No dead settings
Every option does exactly what it says. A toggle either changes the live chat or isn’t shown at all — proven by executing tests, not guesses.
A/B-testable for free
Because every option is an ordinary page field, the built-in experiment engine can split traffic across two configs — centered vs classic, cheap model vs smart — and pick the winner from real conversions. Point ads at each; no new plumbing.
Restrictions that hold
Tool and agent allow-lists are enforced server-side against your stored config. A crafted API call can only narrow what you allowed — never escape it, never escalate.
Workflows on a button
A starter card can launch a workflow — turning “Book a demo” into a real, running flow — reusing the same engine the rest of the platform runs on.
Put it on a page
Open the page editor, drag in the AI Chat block, and set it up in a minute. The chat is live the moment you publish.