Figma adds Codex canvas plugin for editable design outputs

• Figma posted a Codex plugin that turns running code into editable canvas output, with figma's Codex canvas post pointing users to figma's install link.
• Figma's own MCP docs say Codex can both read design context and send live web interfaces back into Figma as editable layers, not flattened screenshots, according to Figma's setup guide and Code to canvas docs.
• OpenAI framed this week as a broader Codex expansion, with six role-specific plugins spanning 62 apps and 110 skills, plus Sites and annotations, per OpenAI's announcement.
• The early creator examples already look less like chat sessions and more like small operating systems, from [danshipper's inbox workflow](danshipper's inbox-zero demo) to aakashgupta's writeup of Abhi Muchhal's PM harness.
• OpenAI also showed Codex moving into native app testing, with [steipete's repost of OpenAIDevs](the Build iOS Apps plugin demo) showing an iOS loop that runs inside Codex.

You can read Figma's February post on Codex to Figma, the later Agents, Meet the Figma Canvas rollout, and the developer docs for writing to canvas and code to canvas. OpenAI's side of the story lives in its Codex for every role, tool, and workflow post, while creators kept stress-testing the idea with self-editing Sites, illustration skills, and inbox agents.

The immediate news is simple: Figma has turned the Codex-to-canvas flow into a plugin install, and the pitch is blunt. Take code, land it on the canvas, then let the rest of the team edit it in Figma.

That is a meaningful shift in ownership. Code-first tools usually make the designer the reviewer of a screenshot or a hosted prototype. Figma is trying to make the canvas the editable handoff layer again.

Figma has been building toward this for months. In its Codex to Figma blog post, the company said teams could bring real running interfaces into Figma to explore and refine them together.

Figma's docs are unusually explicit here. The setup guide says Codex can write directly to the canvas, create or update native Figma content, and send live web interfaces to Figma as editable layers.

The Code to canvas docs break that into two concrete behaviors:

• capture a single rendered screen or an entire flow from the browser
• convert those captures into standard Figma design layers
• reorganize, duplicate, annotate, and refine the result on canvas
• use that flow view to spot structural issues that are harder to catch in a single running state

That distinction matters because Figma is separating two related workflows. Write to canvas is agent-generated Figma structure from prompts and design-system context. Code to canvas is the reverse path, pulling a live interface back into Figma as editable frames.

OpenAI's announcement makes the ecosystem angle clearer than Figma's tweet does. In Codex for every role, tool, and workflow, OpenAI said it shipped six role-specific plugins, bundling 62 apps and 110 skills for non-coding work as well as software work.

One of those bundles is a product design plugin, and the same announcement pairs plugins with two other surfaces: annotations for in-place editing, and Sites for shareable hosted apps. The interesting part is how neatly the pieces line up: generate or update a web app in Codex, capture it into Figma, then push the design conversation back onto a canvas the whole team already knows.

[danshipper's interview with Matt Colyer](danshipper on Matt Colyer and Figma MCP) described the same two-way loop from the Figma side: reconstruct a live web page on the canvas, or hand a Figma design to an agent so it can make changes via pull request.

The bigger story is that Codex is picking up app-shaped workflows faster than most design tools are picking up design-shaped ones. Dan Shipper showed a Codex-native inbox system that turns messages into review cards and draft actions, while Aakash Gupta described Abhi Muchhal's OpenAI setup as three automations running before the workday starts, plus prototypes that replace a PRD with a working artifact.

Those two examples share the same pattern:

• state lives somewhere persistent
• the agent has bounded actions and permissions
• the human reviews checkpoints instead of rebuilding the work by hand
• the interface is a custom app or dashboard, not a long chat transcript

That helps explain why Figma's plugin matters to creative teams. Once Codex outputs start looking like durable tools, getting them onto a collaborative canvas becomes a workflow problem, not a demo trick.

The creator examples are already sprawling into tiny vertical products. Meng To built a screen recording tool in Codex to avoid bouncing between Screen Studio and CapCut for multiple aspect ratios, while bentossell spun up a text editing tool focused just on copy work.

Other creators are using Codex as a skill runner rather than an app generator. venturetwins' illustration thread surfaced a skill that turns long text into hand-drawn explainer graphics, and the linked GitHub repo shows a rigid shot-list workflow with style constraints, asset paths, and QA checks. That is much closer to a repeatable production recipe than a one-off prompt.

Not every result looks polished. petergyang on slide quality argued that Codex still trails Claude on one-shot frontend aesthetics, even while petergyang on integrations and skills said the payoff comes after spending a day wiring skills and integrations together.

The other reveal this week was on the native side. The OpenAIDevs clip reposted by steipete says the Build iOS Apps plugin lets Codex view and test an iOS app in its in-app browser, open SwiftUI previews, and hot reload edits without leaving Codex.

That pushes the same idea past web mockups. If Codex can test a live iOS surface, Figma can absorb a live web surface as editable frames, and creators are already building custom tools around persistent skills, then the center of gravity moves away from chat and toward loops that keep state, run checks, and hand off artifacts people can actually edit.

Even the rough edges point in that direction. LLMJunky's Linux remote control demo shows the desktop app picking up remote control on Linux, which is a small but concrete sign that Codex is still expanding the surfaces where those loops can run.