jtmchorse/tokensmith

Tokensmith is a small server built on the Model Context Protocol, or MCP, a way for AI coding assistants like Claude Code, Claude Desktop, or Cursor to call outside tools while they work. Its job is to make an AI coding agent aware of a project's design system, meaning the official set of colors, spacing, and other visual values a company has agreed on, instead of the agent guessing a shade of blue from memory when it writes interface code. Design systems are usually stored in a standard format called DTCG, short for Design Tokens Community Group, which is basically a structured file listing named values like the primary action color and what it should equal. Tokensmith reads one of these files and offers three main tools over MCP. One lists all the tokens, one resolves a single named token to its real value while following any chains where one token points to another, and one scans a piece of CSS or JSX code to flag any hardcoded colors or sizes that are close to, or exactly match, a token but were typed in by hand instead of referencing it properly. The project is honest that it cannot force an AI agent to actually call these tools while generating code. Instead, the intended workflow is that the agent writes code, a person or process runs the scanning tool against it, and any drifted values get fixed and swapped for the correct token reference. The maintainer includes a demo comparing a hand guessed version of a settings card component against one built using real resolved token values, showing how a wrong guess can look almost identical to the correct color yet still be technically off system. It requires Node version 20 or newer, installs through npm, and is aimed at developers and teams who already maintain a formal design system and want their AI coding tools to respect it rather than reinvent colors from scratch.