Sheets canvas — build mini-apps with Gemini in Google Sheets

Google's announcement (2026-08-13) of Sheets canvas: you describe an interface in plain language in the Ask Gemini side panel, and Gemini builds an interactive "mini-app" over the data already in your spreadsheet. Google's own description is "a dynamic, read-write layer that sits directly on top of your spreadsheet data."

T2, because this is the vendor announcing its own product. It is primary on what shipped and silent on everything a reader of this spoke would want — no model named, no description of what the generated artifact actually is, no limits, no failure cases, no numbers of any kind.

What a user does

Open a spreadsheet, pick Create canvas in the Ask Gemini panel, and describe what you want — Google's example prompt is "create a visual study tracker." Gemini generates the layout, and you refine it by prompting again, editing "layout, design, functionality, and more."

The three worked examples are a study tracker built from a list of assignments, a fantasy-football dashboard over player stats, and a seating chart built from RSVP data with drag-and-drop rearrangement. Domestic, small, personal. Not one of them is the enterprise workflow the Workspace tier list implies.

Why it belongs in this spoke

This is generative-ui — a model emitting an interface rather than text — and it lands as a third distinct strategy beside the two the corpus already holds.

arrow-js generates real framework code and contains it in a WASM sandbox. openui generates a restricted DSL from a developer-defined component set. Sheets canvas generates something Google never names, inside Google's own application, where the containment question does not arise for the user because there is no code surface to escape from. Three answers to what does the model actually emit: real code with bounded reach, a bounded vocabulary, and — here — an artifact the vendor declines to describe.

The property that is new is not the generation. It is the binding. Canvas is bidirectional and persistent: "always in sync" with the sheet beneath it, and "fully collaborative" through ordinary sheet sharing. arrow-js and openui both render output that an agent produces for a viewer; this one produces a durable, multi-user, read-write surface over a live datastore that other people are editing at the same time. Generated UI stops being a rendering of an answer and becomes an application with state.

Nobody has said what happens when the generated app and the underlying data disagree — a column renamed, a row deleted, two people prompting the same canvas in different directions. The announcement's "always in sync" asserts the easy direction and skips the hard one.

No code required, and no code available

"No coding required" is the headline claim, and its converse is unstated: no code is offered either. There is no export, no inspection, no way to see what Gemini built. Compare arrow-js — plain TypeScript, no build step, explicitly so a model emits something a human can read — where legibility is the design goal. Canvas takes the opposite position, and for the audience it names (people who don't write code) that is coherent. It does mean the artifact cannot be reviewed, versioned, diffed, or moved, which is the whole of what this spoke usually means by containment.

Availability

Live globally in English for Google AI Pro and Ultra subscribers. Rolling out from the announcement date to Workspace Business and Enterprise Standard and Plus, and to Google AI Pro for Education. A paid feature on every tier listed.

Related

generative-ui · arrow-js · openui · google · gemini · gemini-managed-agents-hooks · agent-guardrails

Cross-wiki: the framework counterpart of this thread is ../ui-frameworks-wiki's agent-ready-ui — UI toolkits shaped so agents can build with them. Canvas is the other side of that seam: no toolkit is exposed at all.