Design tools spent the week exposing their object models to agents and turning AI output into editable software.
The week in design
The clearest directional move this week was design tools handing their internal object models to agents. Webflow shipped MCP 2.0 and AEO agents that write and draft the copy, not just report on it; ComfyUI made its node graph callable by coding agents through Comfy MCP and put a paid team tier on top; VEED released OpenEdit, an agent-native tool that renders finished clips locally; and Oxygen opened its WordPress builder so an external agent can write pages, Components, selectors and CSS variables directly. Different products, one shape: the canvas stops being the only way in, and the model gets an addressable surface behind it.
The second move is what the AI actually produces. Generation is shifting from a flat artifact toward editable, operable output — layers a designer still owns, flows a developer can build on, assets that get localised in place rather than re-exported. Air turned Canvas workflows into reusable skills and added in-image translation across 33 languages; UXPin shipped Wire to turn designs into working React apps; Linearity and Descript both pulled generation inside the editor on terms that keep the result editable. The theme underneath both moves is the same: the tool is being rebuilt around who — or what — operates it.
Leaders
Air had the sector's busiest week with two sparks against a velocity of 8.8. Turning any completed Canvas workflow into a named, re-invokable skill is the pivot: it makes the agent cumulative rather than episodic, so a team's repeated edits become part of the workspace. In-image translation across 33 languages, with masking for areas to leave untouched, removes the source-file dependency that made localisation a designer's job.
Webflow split its changelog into a visual-builder stream and an AI stream, and the AI stream is where direction is set. MCP 2.0 frames agents operating on production sites with context, control and governance; the AEO agents move from measurement to action, recommending content and drafting the copy meant to get a site cited by ChatGPT, Claude, Gemini and Perplexity. The two look set to converge on agents that read visibility data and publish the fix.
ComfyUI absorbed four frontier models in five days, but the releases that matter are up the stack. Comfy MCP makes the node ecosystem callable by coding agents, and Comfy For Teams puts a whole team on one shared node graph and dependency set — the first release sold explicitly as a fix for ComfyUI's most persistent operational problem, that two machines rarely run the same graph.
VEED is productizing its renderer. OpenEdit is an agent-native tool that transcribes, styles captions and renders on the user's own machine through VEED Engine, its HTML-to-video renderer. Read alongside the earlier Subtitle API, ChatGPT app and Make connector, the editor is becoming one front end among several on the engine rather than the product itself.
shadcn/ui kept a near-weekly cadence and stopped being a single component library. React Aria joining Base UI as a selectable component base changes what shadcn is: a distribution layer that can sit over several headless primitive foundations rather than one, with server-side registry search and the typeset and helpers packages extending the same own-the-code model outward.
Wildcards
Lucide went the other way from the AI crowd, and its move is structural. Folding the @lucide/lab incubator into the main monorepo — dropping 16 icons as they graduate into core — plus SHA-pinned CI actions and explicit workflow permissions reads as governance and supply-chain posture for a package with a large install base, not feature growth.
Shotcut restored external-monitor output on a second system display, three years after a UI-library migration removed it. It required reworking the video plumbing, which makes it a structural change rather than a menu fix — an open-source project spending a release undoing its own most-missed regression.
Themes that compounded
- Agent-addressable surfaces via MCP showed up across Webflow, ComfyUI, Oxygen and ProtoPie, which shipped ProtoPie MCP in beta.
- Generation is being kept editable: Air skills, Linearity layer-preserving output, and Descript regional inpainting all treat the model result as a starting document, not a final image.
- Prototype-to-software crossed from mockup to running code, with UXPin Wire exporting React apps.
- Decisions are being taken off the user: Kittl made Agentic AI the default way to create, removing the prompt, model and settings choices.
- Distribution over destination: VEED and shadcn/ui both moved value from the app to what other software can call or install.
Watch this week
Watch whether the agent surfaces gain team- and workspace-level scope. Air skills are recorded per person but are far more valuable shared; ComfyUI has attached a paid team tier to its MCP surface; and Webflow framed MCP 2.0 around governance. The next visible step is a saved skill, workflow or agent permission that belongs to the team rather than the individual — plus, from UXPin's in-editor credit sales, the first sign of usage-based metering following the generation volume.