GitHub Copilot gains canvases, full-stack mini-apps that sidestep the chat interface

GitHub has introduced canvases, a new interface for its Copilot app that positions small, purpose-built applications as an alternative to the chat box that has so far dominated AI-assisted development. According to the GitHub Blog, a canvas is a full-stack application that runs inside the Copilot app without the surrounding browser interface, and its server component communicates in both directions with the Copilot agent.

The post, titled "When chat is the wrong UI" and published on September 24, argues that chat became the default surface for language models largely by accident: it was the first thing users connected with, and it works as a catch-all when nobody can predict what people will try to do with AI. Once you know your own task, the author contends, a plain text field is often a poor fit. The piece cites academic Steven Pinker, who has suggested that AI's first large-scale implementation, a chatbot, amounted to a gimmick, and that the technology's real promise lies in task-oriented applications.

Full-stack apps, not web pages

Canvases are more than rendered web pages. Because each one includes a server component, a canvas can call third-party APIs and execute code locally on the developer's machine. The agent can send messages to that server and receive responses, so the canvas behaves like an ordinary program that happens to have an AI agent attached.

Creating one is reportedly as simple as describing it. The post shows a prompt requesting a Connect 4 game to demonstrate user interaction, canvas-to-agent communication, and agent control of the canvas. The Copilot app already understands what a canvas is, so no extra explanation is required. In the author's demo he loses to GPT-5.6 Sol running with high reasoning enabled and beats GPT-5.6 Luna with reasoning disabled, a result he blames on the game's difficulty.

Build the tool once

The post's central economic argument concerns tokens. When chat is the only interface, users are nudged into delegating every action to the agent, which the author sees as a needless drain on tokens. It is nearly always better, he argues, to have the agent build a tool whose future interactions cost nothing than to treat the model itself as the tool — his example is repeatedly asking a model to stage and commit changes.

The sample canvases range from playful to practical. A Winget canvas browses the package registry and manages local installs and uninstalls, with no AI involved in daily use. A SQLite canvas offers a database front-end complete with intellisense, so the developer can write queries directly. Another resurrects a Windows Live Writer-style editor for drafting Jekyll posts in something richer than plain Markdown.

Taking yourself out of the loop

The more ambitious use case is workflow automation. The author outlines his process for working with agents — research, prototype, plan, implement, iterate, finalize — and observes that most of it does not require a human at the keyboard. An agent can research and generate prototypes, then signal when a review is needed. The obstacle is that this kind of orchestration is hard to arrange when a chat box is the only means of interaction. A canvas can encode the workflow and let the developer stay in the loop as much or as little as they choose.

Maturity varies by task. The SQLite canvas could be generated in a single shot, while the workflow automation canvas took most of a day to get right.

Why it matters

The argument carries weight coming from GitHub, the company behind one of the most visible AI coding products. Canvases reframe the agent's job from answering an endless stream of prompts to producing durable, reusable interfaces, which changes both the token economics and the ergonomics of AI-assisted work. It also points at a broader shift: agents that build the tools their users need, rather than agents that are the tool. The caveats are real — the examples skew toward demonstrations, and non-trivial canvases still require meaningful iteration — but the direction is unmistakably away from chat as the universal interface for AI.