Some answers are easier to see than to read: a comparison, a before and after, a breakdown, a flow. With Visual Canvas on, an agent can answer with a small designed panel right in the conversation, with charts, cards, tables and diagrams, instead of a wall of text or ASCII art. Plain text stays the default. The agent draws a canvas only when the shape of the answer carries the meaning.
Asked to compare three backoff strategies. The answer is one canvas and a sentence of recommendation under it.
Turn it on
Open Settings (⌘,), then Agent Features under Machines, and pick the computer your agents run on.
Switch on Visual Canvas, under Showing the work.
Agents started from then on know how to draw one. Click How it works on the card to see what it adds and a request to try.
The Visual Canvas detail. The switch at the top right turns it on for the selected computer.
Ask for it in plain words
Say what you want to see. Mentioning a visual helps, but agents also choose one on their own when it fits:
"Compare the three backoff options as a visual."
"Show me webhook failures by hour, before and after the fix."
"Draw the flow of a webhook from the provider to our database."
"Lay out the open pull requests by status."
What a canvas looks like
Real canvases, drawn by the same component the app uses. Pick an example or a theme: a canvas follows whichever theme you use.
It matches the app. A canvas uses the app's fonts and colors, and follows you when you switch between light and dark.
Small interactions work. Tabs, tables you sort by clicking a header, tooltips on hover, and charts with hover labels.
Pictures show in the message. When an agent wants to show you a screenshot or a chart it rendered, it uploads the image, so you see the picture rather than a path on its machine that your browser can't open.
The canvas header
The header shows the canvas's title and three buttons: Show source to see the HTML behind it, Copy HTML, and Fullscreen. Press Esc to leave fullscreen. A tall canvas folds behind a Show all button so it never swallows the conversation.
Where canvases show up
A canvas is part of the message. It stays in the transcript, appears for teammates reading the session, and shows in shared links and in decisions, where an agent can lay options side by side.
On the iPhone a canvas shows as a card with its title. Tap it to open it full screen, where you can also view its source or copy it.
On the phone, tap the card to open the canvas full screen.
What it will and won't do
Conversations sync across your team, so a canvas is something other people see without having written it. It is built to be safe to look at:
No code runs. Scripts, forms, embedded frames and anything that reacts to clicks are removed before the canvas appears. Tabs, sorting and charts are codecast's own, not the agent's.
Nothing phones home. Images from other websites are dropped, so a canvas can't track who opened it. Images an agent uploaded to codecast, and images drawn inside the canvas, show normally.
It stays in its box. A canvas can't restyle the app around it, and links in it open in a new tab (links to codecast's own sessions and tasks open in the app).
Canvas or page?
A canvas lives inside the conversation. When the result is something you will send to other people by link, a report for a stakeholder or a dashboard that runs its own code, ask for a published page instead.