Explore: MCP server exposing live token/class data to AI coding assistants

[jackgranatowski]

Idea

Build an MCP (Model Context Protocol) server that exposes SLASHED's generated data — token-registry.generated.json, classes.generated.json, api-index.generated.json — as live tools/resources (e.g. search_tokens, get_class_info, suggest_recipe_for(intent)) that AI coding assistants (Claude Code, Cursor, Windsurf, etc.) can query directly.

Why

A static context file (see companion "llms.txt" issue) goes stale between releases and has to be manually re-pulled into a user's assistant config. A live MCP server always reflects the currently installed/pinned SLASHED version and lets an assistant look up exact token/class definitions on demand instead of relying on a context dump, which scales better as the token set grows.

Scope notes

• New package, likely packages/mcp-server or similar — doesn't touch existing core//configurator/ code.
• Should read from the same generated JSON artifacts the configurator already consumes, not duplicate the data pipeline.
• Needs a decision on distribution: npm package the user runs locally (npx @slashed/mcp-server) is the simplest starting point vs. a hosted server.
• Larger/riskier follow-on (not in scope here): a natural-language "suggest classes for this intent" feature built on top, which would need a design decision on bring-your-own-API-key vs. hosted inference — track separately if pursued.