Figma's Dev Mode MCP Server brings the design file straight into the IDE
The Dev Mode MCP Server connects agents like those in Cursor, Copilot, and Claude Code directly to Figma files, trading manual exports and screenshots for automatic reading of components, variables, and Auto Layout, but it only works well if the design library is organized the right way.

- Reading
- 6 min
- Tables
- 1
The Dev Mode MCP Server connects agents like those in Cursor, Copilot, and Claude Code directly to Figma files, trading manual exports and screenshots for automatic reading of components, variables, and Auto Layout, but it only works well if the design library is organized the right way.
Figma published the official guide to the Dev Mode MCP Server, a document that describes how the design editor started exposing its files to AI agents running inside tools like Cursor, GitHub Copilot, Claude Code, and other clients compatible with the Model Context Protocol (MCP). In practice, the server is a bridge: instead of the developer opening Figma, looking at the frame, noting down spacing and color, and then trying to replicate that in code, the AI agent reads the file's structure directly and uses that data as context to generate the interface.
MCP (Model Context Protocol) is the standard that enables this conversation: an AI client (the IDE) talks to a server (Figma) using a common interface to fetch structured data. Figma lists in its catalog which clients already support the integration, and the list includes names like Android Studio, VS Code, Claude Desktop, Amazon Q, Replit, and JetBrains, among others, each with a different level of support for features like canvas writing and skills.
Two servers, two levels of access
There are two ways to connect the MCP Server to the code editor, and the choice between them isn't just technical: it determines which features are available and under which Figma plan.
| Feature | Remote server | Desktop server |
|---|---|---|
| Where it runs | Endpoint hosted by Figma (mcp.figma.com/mcp) | Locally, via the Figma desktop app |
| Plan required | All plans | Dev or Full seat on a Professional, Education, Organization, Government, or Enterprise plan |
| Canvas writing | Yes | No (the feature requires the remote server) |
| Figma's recommendation | Best option for most | Specific organization/enterprise cases |
The remote server is the path Figma itself recommends, since it offers the broadest set of features without depending on the desktop app. The desktop server, in turn, serves more specific needs of companies that already operate with paid seats and need to run the integration locally, within the organization's network perimeter.
What the agent can do with the file
The most common flow, described in the guide, is simple to reproduce: the user selects a frame or layer in Figma, clicks "Copy link to selection," and pastes that link into the conversation with the agent in the IDE. The client doesn't open the link; it extracts the node ID that the MCP Server uses to identify exactly which object to fetch and return as context.
From that node, the agent can:
- Generate code from selected frames, useful for product teams building new flows or iterating on existing features.
- Extract design context: variables, components, and layout data, especially valuable for those working with design systems and component-based flows.
- Pull resources from FigJam, bringing diagrams and architecture maps from early ideation phases into the development flow.
- Gather resources from Make files, helping the transition from prototype to production.
- Write back to the canvas in Figma Design, FigJam, and Figma Slides, creating and updating frames, components, variables, and Auto Layout from instructions in the IDE.
This last capability, writing to the canvas, is the one that most changes the direction of the flow: until now, information went from design to code; now it can also go back. The feature requires the remote server and a Full seat to write outside personal drafts, is free during the beta period, and, according to the guide itself, will become a pay-per-use feature in the future.
Why library organization became a technical requirement
Here's the point that matters most for those who design, not just those who code. An AI agent doesn't "see" a frame the way a designer does: it reads the layer tree, the names given to each element, the linked variables, and the Auto Layout structure. If a button is named Rectangle 47 instead of Button/Primary, the agent generates code with that meaningless name, or simply misses the intent.
This changes the priority of three practices that, until recently, were seen as optional good organizational practice:
- Consistent naming of layers and components, because the name is the only semantic label the agent receives about what that element represents.
- Well-applied Auto Layout, because it's the source of spacing, alignment, and responsive behavior information that the generated code tries to reproduce.
- Centralized component and variable library, because the guide is explicit in saying that, with the right skills, agents use "the design system as the source of truth" ("your design system as the source of truth," in the original English text) to create and update content.
In practice, teams that already maintained design system discipline gain a direct advantage: the same hygiene work that helped human designers reuse components now also improves the quality of the code the AI generates. Teams without that discipline will find the MCP Server returning inconsistent results, not because the tool failed, but because the input material was ambiguous.
Code Connect and Figma Specs: the consistency layers
The guide points to Code Connect as the mechanism that keeps generated code aligned with what already exists in the repository: instead of the agent inventing a component from scratch, it reuses the actual components from the codebase, reducing duplication and visual drift between what's in Figma and what's in production. For those who maintain a mature design system, this is what differentiates generating "a similar-looking button" from generating "the Button the team already uses."
The guide also mentions the figma-spec skill, still in open beta, which lets the agent read, create, and update team specs, sharing requirements before building starts and recording decisions captured during the work back into the spec. It's an attempt to solve a classic handoff problem: requirements documentation that becomes outdated as soon as the code starts diverging from the design.
Where the feature still falls short
It's worth naming the limits, because they directly affect the decision to adopt or not:
- Canvas writing and generating design from live UI require the remote server and, in the case of writing, a Full seat, which excludes teams on the free plan only or with view-only seats.
- Support varies by IDE client: Figma's compatibility table shows variation in which clients support the desktop server, remote server, canvas writing, and skills, so it's worth checking the catalog before promising the full flow to the team.
- The writing feature is free only during the beta, and will become pay-per-use, which changes the cost calculation for teams that depend on it daily.
- Custom fonts only work if they're already loaded into the Figma account; the server doesn't import fonts from outside.
For disposable prototypes, quick exploration files, or teams without an established naming convention, the MCP Server's gain is marginal: the agent will generate code, but without the quality that comes from a clean design system behind it. In those cases, traditional manual export can still be more predictable than relying on an agent interpreting a messy layer tree.
What changes in the daily work of those who design
The practical effect, for designers and for those leading product teams, is that the quality of naming and Auto Layout structure has stopped being just a matter of internal organization and has literally become the interface AI uses to understand the work. It's worth reviewing component libraries through that lens before betting the entire handoff flow on the MCP Server: the bottleneck, when it appears, is usually in the file, not in the AI.
Translated from the Brazilian Portuguese original · Read the original
Which Qualitative Research Method to Use at Each Product Stage
A practical guide based on material from the Nielsen Norman Group for deciding between usability testing, interviews, field studies, and diary studies, with what changes when applying this with a real team in Brazil.





