Skip to content
All posts

MCP servers for design work in 2026: what each one lets an agent do

An MCP server lets an AI agent, such as Claude Code, Cursor or Codex, work with a design tool directly. What the agent may do differs a lot from tool to tool. Some servers give an agent the design as context, some let it write to the canvas, and free plans meter that work in different ways. Here are eight design tools side by side, in no order of preference.

We make Brilliant, which is one of the eight, so every row below carries the date we checked it and links the vendor. A cell that says not checked is one we have not verified from the vendor's own pages yet. It is not a no.

The table

ServerReadWriteWhere the output livesFree limitDate checked
FigmaYes, design context for agentsYes, free during its beta, then usage-basedIn the Figma cloud20 MCP tool calls a month on Starter2026-11-06
PenpotNot checkedYes, to the focused page, from a remote or local serverPenpot's open file format, in Penpot cloud or on your own serverProfessional plan at $0. MCP limit not checked2026-11-06
CanvaNot checkedNot checkedNot checkedNot checkedNot checked
FramerNot checkedNot checkedYour site, in your Framer projectThe free plan includes the MCP server, free during its preview2026-11-06
PaperYesYesPaper's design space, as HTML and CSS, on the web and desktop100 MCP tool calls a week2026-11-06
PencilNot checkedYes, over a local MCP serverA .pen file (JSON) kept in git beside your code5 days a month designing with an agent2026-11-11
SubframeNot checkedNot checkedNot checkedNot checkedNot checked
BrilliantYes, canvases, design systems and your selectionYes, create, edit and run canvas commands, each with undoYour Brilliant canvas, as Blueprint plain text, on your disk with the desktop appUnlimited canvas work by your own agents. Images use your Brilliant AI usage2026-11-18

Pencil is now called pen.dev, and its MCP server is still named pencil.

What the columns mean

Read means the agent can read the design: its layers, text and structure, as context for code or for its next change. Write means the agent can create or change elements on the canvas itself, so the design is something you can open and edit afterwards. Where the output lives is where a design the agent made ends up. Free limit is what the free plan allows an agent, in the vendor's own units.

Questions

Which design tools have an MCP server an agent can write to?

Of the eight, the pages we checked say Figma, Penpot, Paper, Pencil and Brilliant let an agent write to the canvas. Figma's canvas writes are free during its beta, then usage-based. Canva, Framer and Subframe were not checked for writing.

Which design MCP servers are free?

Each free plan meters agent work in its own unit. Figma allows 20 MCP tool calls a month on Starter, Paper 100 MCP tool calls a week, Pencil 5 days a month designing with an agent, and Brilliant unlimited canvas work by your own agents. Framer's MCP server is free during its preview. Penpot's Professional plan is $0, and its MCP limit was not checked.

Can Claude Code, Cursor or Codex use these servers?

Each vendor documents its own install route. Pencil lists Claude Code, Codex and Gemini CLI among its MCP clients. Brilliant connects to any of them with one command, npx -y @brilliant-hq/mcp install cursor (or claude-code, codex and more), and needs Node 20 or later and no API key.

How current is this table?

Each row carries the date we checked it. Prices and limits move, so check the vendor's page before you choose, and tell us if something has changed.

Sources and dates

Related: Design canvases for coding agents, compared, and the agent pages for Claude Code, Cursor and Codex.