Why it mattersThe handoff to a coding agent becomes a described design rather than a screenshot, which moves the vague part of the work from the agent's guesswork to a decision someone made on purpose.
M3E Canvas is a browser tool for sketching Material 3 Expressive screens and turning them into a written prompt for a coding agent. The GitHub API returned 4,065 stars, 346 forks and an MIT licence when the repository was checked on 6 September 2026, against a creation date of 2 September, so that traction arrived in four days.
The repository is TypeScript, at 831,411 bytes of it against 4,781 bytes of CSS. Five people have commits, with the owner at 51 of them. The last push was earlier on the same day the numbers above were read.
What it produces
The README describes the loop: drag components onto a canvas, link screens with tap and swipe navigation, preview the result by tapping through, then export. The export is the part that matters. It produces a natural language prompt describing the design, in Japanese, English, Chinese or Korean, alongside PNG screenshots and a shareable link.
It ships around 25 component types, and the README lists them: buttons, icon buttons, FABs, split buttons, FAB menus, chips, app bars, navigation bars, floating toolbars, tabs, search bars, cards, lists, dialogs, snackbars, text fields, dropdowns, switches, checkboxes, radio buttons, sliders, text, images, camera and map placeholders, badges, boxes and dividers. Screens come in two sizes, 412 by 892 for phone and 1280 by 800 for desktop, switchable per screen.
The tool writes no code itself. It hands the prompt to Claude, Codex or whatever agent you already use, and that agent writes the implementation.
What it costs to adopt
Close to nothing, which is the unusual part. It is a static export that runs in the browser with no server behind it, and work saves to localStorage. An API key is optional and only powers the AI-assisted component descriptions, so you can bring your own from OpenAI, Anthropic, Google or DeepSeek or skip that feature entirely. The full multi-screen editor wants a desktop browser; the README calls the mobile experience phone-friendly rather than complete.
The credits are small and checkable: the loading indicator comes from Material Components Android under Apache-2.0, and the icons are Material Symbols from Google Fonts under the same licence.
Most of what is written down about this project describes the design side, and there is no published measurement of whether an agent given one of these prompts produces better output than an agent given a screenshot. Three issues were open at the time of the check, and there are no tagged releases yet.
The interesting move is where the ambiguity goes. Handing a coding agent a picture of a screen leaves it guessing at the parts a picture cannot carry: which button is the primary action, what happens on tap, which of two similar containers is a card and which is a box. A prompt naming twenty-five component types by their Material 3 names removes that guessing, because someone chose the component before the agent ever saw it. That is a smaller claim than it first appears, and it is the honest one: the tool does not make the agent better, it makes the brief less vague. Whether four days of stars turns into people actually shipping from these prompts is the question nothing has answered yet.
Source
lnkiai/m3e-canvas, GitHub. Star count, fork count, licence, language breakdown, contributor counts and creation date read from the GitHub API on 6 September 2026.
Source: GitHub
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually shipping with them. Short, and only when there is something worth reading.