Agent access: an MCP server for your design system

A file is read whole, and it says only what it said when it was written. A server answers the question asked, from the current version, and it can act. That last part is what files cannot copy. GitHub Primer’s server includes a tool that lints CSS. Shopify’s validates the code of UI extensions. shadcn/ui’s installs components into the project. An agent with a server can check its own work before anyone else sees it.

• Tools are functions the agent decides to call, each with typed inputs described in JSON Schema: find a component, get its props, search the design tokens, lint a snippet.
• Resources are data the agent can read for context: a docs page, the token file, a list of icons.
• Prompts are ready-made instructions that tell the model how to use the tools and resources for a common job.

A design system server is mostly tools and resources, the first for questions and checks, the second for reading. Prompts help when the same task, such as building a form, comes up often.

Of the 12 public systems, 8 do, each from its own team: IBM Carbon (hosted, with an IBMid sign-in), Atlassian (hosted or from npm), Adobe Spectrum 2, Ant Design (from its CLI), shadcn/ui, Chakra UI, GitHub Primer and Shopify Polaris (inside Shopify’s server for all its platform). Microsoft Fluent 2, Google Material 3 and the GOV.UK Design System have none; a Fluent maintainer wrote in February 2026 that its server has not been released outside Microsoft. Salesforce Lightning had one inside the Salesforce DX server until Salesforce retired it on November 2, 2026. Each of those 4 loses 2 points on the best AI-ready design systems, ranked.

Not first. A server is software someone has to run and keep in step with every release, and the files come before it: design tokens as data, docs agents can read and written guidance. Build the server once one of these is true:

• The catalog outgrew one file. Agents need to search hundreds of components or icons instead of loading them all.
• You want checks in the loop. A lint tool lets the agent test its own output against your rules.
• Installing is part of the job. The agent should add components to a project the way your team does.

Start small and read-only: list the components, return one component with its props, usage notes and an example, search the design tokens by name or job, and return the guidance for a kind of screen. Add one checking tool next. Generate every answer from the same source as your docs and package, so the server cannot drift from them. A custom build can include a server like this; Work with Simbolik (opens in a new tab) if yours needs one.