MCP Server
Bezel's MCP server puts your design tokens inside your AI editor. Instead of pasting a token file into chat and hoping it stays current, your assistant queries the live project: the same values you see in the editor, including every theme and export name.
Ask for "a pricing card using our tokens" and the agent looks up what you actually have, then writes CSS referencing your real variables rather than inventing hex codes.
Available on Pro and Enterprise plans.
What your assistant can do
| Read your tokens | List projects, inspect structure, query by group or type, resolve a single path with its aliases |
| Work with themes | Discover contexts, validate a context name, pull values for a specific theme |
| Generate code | Export CSS variables or JS/TS theme objects, or framework config for Tailwind, shadcn/ui, and Mantine |
| Check its work | WCAG contrast checks, color conversions and ramps, unit math, and a design review against your own tokens |
| Recall context | Search the project's AI chat history to see decisions already made |
Full list in the Tool Reference.
What you can ask
Set up a repo (how it works):
Set up Bezel in this project.
Set up this project to use my Bezel project named Acme Design System.
Configure Bezel Kit in this repo and generate my variables.css.
Keep tokens in sync (how it works):
Update my design tokens.
Sync my Bezel tokens.
Pull the latest tokens from Bezel.
Update my tokens to version 1.4.0.
Manifest first
The server is built around one idea worth knowing, because it shapes how your assistant behaves.
Token projects get big. Dumping an entire project into a chat window burns context and often exceeds it, leaving the agent with a truncated blob it can't use. So the server leads with a manifest: group names, token counts, types, and available contexts, but no values. The agent reads that first, decides what it needs, and pulls only those tokens.
In practice you'll see your assistant make two or three small calls instead of one huge one. That's the design working.
The one place the whole file is the right answer is your repo, and get-design-tokens-file writes it straight to design-tokens.json rather than into the chat.
Getting started
- Install Server — one command, OAuth, no API keys
- Try a workflow — prompts that get useful results
- Browse the tools — everything the server exposes
To wire tokens into a codebase as real CSS files, see the Quick Start. The MCP server is how the tokens get there.