When chat is the wrong UI

What it claims

A canvas in the GitHub Copilot app is a small full-stack app with a server side. The Copilot agent talks to that server, and the server can talk back. This differs from a chat-generated web page because the canvas can execute code locally, not just call remote APIs. The author's Winget canvas installs and uninstalls local packages, and it contains no AI at all. That is the point he makes about cost: once the agent has built the tool, using it costs no tokens. His example of waste is asking a top-tier model to "stage and commit" through chat.

How to use it

You create a canvas by asking for one in plain language. The app already knows what a canvas is, so you don't describe the architecture. The prompt used for the Connect 4 demo:

Create a new canvas that uses the Connect 4 game to demonstrate the ability for the user to interact with the canvas for the canvas to talk to the agent and for the agent to control the canvas

That prompt asks for both directions: canvas to agent and agent to canvas. Use it as a template when you want the agent to drive the UI as well as respond to it.

Expected effort varies a lot by kind of canvas:

• Simple tool UIs, such as the SQLite query canvas with IntelliSense, can be one-shot from a single prompt.
• Workflow automation canvases take much longer. The author's research → prototype → review canvas "took the better part of a day to get the design and automation right." It lets the agent run research and prototyping on its own and notify him when something is ready for review. He says you can pull yourself out of the loop "as much or as little as you like."

Limits

• The page gives no API or protocol details for agent-to-server messaging. It also doesn't say where canvases are stored, how they are shared, or which models or plans support them.
• The workflow steps appear only in an image that isn't in the text. Treat the specific workflow as unknown, not reproducible.
• Running local code has implications: a canvas can install and uninstall software on your machine. The page does not discuss permissions or sandboxing.
• The author says his workflow is not a recommended template.
• The Connect 4 results are anecdotes, not evaluation. He lost to GPT-5.6 Sol on high reasoning and beat GPT-5.6 Luna with no reasoning.