Figma: Agentic workflows and the MCP | Config 2026

Config 2026 produced a dense schedule of product announcements, but the deep-dive sessions are where the practical substance lives. This particular session focuses on agentic workflows and Figma’s MCP server—a combination that is reshaping how design context reaches external coding agents like Claude Code and Cursor.

The session opens with a framing point that recurs throughout: clarity at the input stage determines output quality more than the model does. When agents receive design context that is vague, inconsistent, or missing token alignment, they produce outputs that require significant human correction. When context is structured—components named consistently, variables tied to CSS tokens, design intent made explicit rather than inferred—the agent’s output converges on what the designer actually intended.

What the MCP server does

Figma’s MCP server acts as a bridge between the design file and any external agent that supports the Model Context Protocol. Rather than exporting assets or copying specs manually, the agent can query the design file directly: what are the current spacing tokens, what component variants exist, what constraints apply to this layout? The session walks through the mechanics of setting up the connection and demonstrates how the agent uses that context in a realistic coding task.

The demonstration is useful precisely because it shows failure modes as well as successes. When the agent is given only a screenshot to work from, it guesses at spacing and color values. When it has access to the Figma file via MCP, it extracts the actual values and produces code that passes design review without a correction round.

Design context as the new spec

A theme that emerges from the session is that the traditional design spec—annotated redlines, handoff documentation—is being replaced by live context that the agent can query dynamically. This is not just a workflow efficiency; it changes what designers need to maintain. Instead of producing a spec document after the design is done, the design file itself needs to be machine-readable throughout the process. Components that are inconsistently named, variables that are not mapped to tokens, or layers that are not properly organized all become agent-blocking problems rather than cosmetic issues.

Who should watch

This session is most useful for designers already using Claude Code or Cursor in their workflow and trying to understand how to make the Figma-to-code loop more reliable. It assumes familiarity with MCP as a concept and does not spend much time on introductory material. It is also worth watching for design leads responsible for maintaining design systems that AI agents will consume—the examples in the session make clear that system quality directly affects agent output quality.

The runtime is around 25 minutes, which is about right for the depth of content. It is the kind of session that rewards a second watch once you have tried setting up an MCP connection yourself.