Figma

Connects Figma files to AI coding agents over MCP, exposing design context and writing content back to the canvas.

Collaborative interface design tool with real-time multiplayer editing, design systems, and Dev Mode.

Figma's 2026 AI report, based on 8,400+ surveys and 639 interviews, finds AI-driven collaboration overtaking solo AI use, with designer/developer role crossover rising sharply and most work shifting onto a shared canvas.

Figma Design added code layers: agent-generated, interactive code blocks embedded on the canvas that can import a GitHub repo or local folder, be edited in place, and push changes back to source.

Figma's advocacy director explains what agentic coding tools actually consume through the Dev Mode MCP server, and how to prepare for it: connect components to real code via Code Connect, define variable code syntax per platform, and organize files so an agent infers the right structure.

Figma has added Gemini 3 Pro to Figma Make as an experimental model, with Gemini 3 Flash alongside it for faster iteration. Figma's design leads walk through trials turning static designs into code-backed prototypes, including physics animation and a Supabase-connected form.

Figma moves Make and its AI features out of beta for every seat tier. The notable parts for builders: design-library import for consistent output, a Supabase connection so generated prototypes can talk to a real backend, and credit limits that vary sharply by seat.

Figma Sites adds code layers: interactive elements backed by custom React, created by converting components, prompting the Figma Make powered chat, or writing in the built-in code composer. Variations can be generated and compared directly on the canvas.

Figma engineers describe how code layers work: React components rendered as canvas layers, an embedded IDE, one-click conversion from design to code using the Figma Make model, and the concurrency model that lets designers and developers edit the same generated code.

Winners of Figma's Make-a-thon walk through how they prompted their way to working prototypes, including prompt counts, structure-first prompting, and treating revert as a cheap rollback when a build breaks. Useful pattern material for anyone building with generative app tools.

Figma explains why it pulled Make Designs: the feature was stock GPT-4o plus Amazon Titan with no fine-tuning, steered by commissioned design-system components and example screens injected as context. Assets added days before launch went unvetted and produced mocks resembling shipped apps.

Figma engineers walk through the infrastructure behind semantic search over designs and components: an embedding model fine-tuned on public Community files, billions of embeddings generated and indexed across every frame, and the cost and indexing tradeoffs that made it shippable.

Figma and ServiceNow shipped an MCP integration that feeds design context straight into ServiceNow's Build Agent, so a design file becomes the prompt for generating an enterprise application. A concrete example of MCP carrying structured design state into a code-generating agent.

Figma shipped an MCP server in beta that feeds design context (components, variables, styles, multi-screen flows) to agentic coding tools. It replaces screenshot-and-hope prompting with structured design data, with remote server support and deeper codebase integration promised next.

Figma launches Make, a prompt-to-app tool that carries a design's structure and metadata into a natural-language prompt loop to produce an interactive, code-backed prototype, initially powered by Claude 3.7 Sonnet.

Figma's designer advocates share eight prompting practices for Figma Make, built on Claude Sonnet 4: front-load detail in the first prompt, apply Anthropic's prompting guidance, and restart fresh rather than patching a drifting generation.

PMs from Figma, Duolingo, Asana, and LinkedIn describe how they decide whether a feature should use AI at all, built around three concrete questions on data leverage, error tolerance, and user trust, before committing engineering time to an AI feature.

Figma's Copy design feature converts any Figma Make AI-generated preview into editable native layers on the canvas, closing the gap between prompted generation and a workable, remixable design file.

Figma's Dylan Field and YC's Garry Tan discuss why current AI stays a 'tool' that widens exploration but hasn't closed the design-quality gap, and why craft and judgment still matter when models make prototyping nearly free.

Figma engineers detail the path from a hackathon autocomplete prototype to production AI search: visual search via screenshot embeddings and semantic search built on Retrieval-Augmented Generation, plus the scoping and testing that got it right.

Figma surveyed 2,500 users for its 2025 AI report: agentic AI is the fastest-growing AI product category, adoption has roughly doubled year over year, and there's a gap between how developers and designers perceive AI feature quality.

Practitioners from Replit, Cline, and Val Town react to Karpathy's 'vibe coding' framing, tracing it from punch cards to AI-driven development and citing Replit's claim that 75% of its customers never write code.

Figma's MCP server now supports remote access, so IDEs, AI coding agents, and browser tools can pull live design context without the desktop app, and it adds new partner integrations including Android Studio, Replit, and Warp.

Figma's CTO argues AI is commoditizing code translation (design to code, language to language) but not the harder work of engineering: choosing problems, building abstractions, and reasoning about systems from first principles.

Figma argues that structured design systems are what let MCP-connected AI agents generate on-brand, accurate code instead of generic UI, and explains how the MCP server can also help kickstart a system that doesn't exist yet.

Design leads from Perplexity, a16z, and Box debate what agentic AI (Operator, Perplexity Assistant) means for UX: minimizing exposed complexity, treating agents as cross-platform interoperable actors rather than chat personas.