Figma Opens Its Design Canvas to AI Agents With Bidirectional MCP Server and a Skills Framework for Team Conventions
Figma's new use_figma MCP tool gives coding agents write access to design files, while a markdown-based Skills framework lets teams encode conventions for agent-driven workflows.
Overview
Figma has opened its design canvas to AI coding agents, launching a beta of its use_figma Model Context Protocol (MCP) tool on March 24, 2026. The tool grants agents direct write access to Figma files, enabling them to create and modify components, apply variables, configure auto layout, and build full screens using a team’s existing design system. The move follows weeks of integration announcements that have progressively turned Figma’s MCP server from a read-only design context provider into a bidirectional bridge between code editors and the canvas.
What We Know
The use_figma MCP tool is compatible with a broad set of AI coding clients, including Claude Code, OpenAI Codex, Cursor, GitHub Copilot CLI, Augment, Factory, Firebender, and Warp. The integration with OpenAI’s Codex was announced by Figma in late February, with the company’s chief design officer Loredana Crisan stating that the partnership lets “teams build on their best ideas — not just their first idea — by combining the best of code with the creativity, collaboration, and craft that comes with Figma’s infinite canvas.”
The bidirectional capability was first surfaced on March 6, when GitHub announced that Copilot users in VS Code could pull design context from Figma into code and push rendered UI back to the canvas as editable frames. That update established a continuous design-to-code-to-design loop: developers generate code from a Figma design, render the result, send the rendered output back as editable layers, and iterate without leaving their editor.
Alongside the write-to-canvas beta, Figma introduced a Skills framework — a system of markdown files that encode team-specific conventions and design workflows. Skills define how an agent should behave on the canvas, specifying which components to use, what naming conventions to follow, and which design tokens to apply. Nine community-built skills shipped at launch, including /figma-generate-library for creating components from codebases, /apply-design-system for connecting designs to existing system components, and /create-voice for generating accessibility specifications. Anyone can author a skill without writing code.
Enterprise adoption is already underway. Uber has built uSpec, an agentic system that uses Figma’s MCP to automate the creation of component design specifications, as detailed by InfoQ. uSpec connects an AI agent running in the Cursor IDE to a local Figma Desktop session via a WebSocket bridge, enabling the agent to crawl the component tree and extract design tokens, variant axes, and accessibility semantics. A notable architectural choice is that all processing remains local — no proprietary design data leaves the network.
What We Don’t Know
Figma has confirmed that write-to-canvas will eventually move to a usage-based pricing model, but has not disclosed rates or timelines for the transition out of free beta. It remains unclear how the pricing will interact with existing seat-based plans, particularly for Dev seats that currently have read-only access outside drafts.
Early testing has revealed non-deterministic behavior: the same prompt can produce different results across runs, and agents tend to default to hardcoded values for colors and variables unless explicitly instructed otherwise. Whether the Skills framework can reliably close this consistency gap at scale remains to be seen.
The broader question is whether design systems — historically built for human consumption — are structured well enough for agents to navigate effectively. Figma’s own documentation acknowledges that agents given access to poorly organized design systems will produce correspondingly inconsistent output, suggesting that organizations may need to invest in restructuring their design systems before realizing the full benefits of agent-driven workflows.