Claude Code Plugins

Community-maintained marketplace

Feedback

Ref.tools MCP server for SOTA documentation search (60-95% fewer tokens than alternatives).

Install Skill

1Download skill
2Enable skills in Claude

Open claude.ai/settings/capabilities and find the "Skills" section

3Upload to Claude

Click "Upload skill" and select the downloaded ZIP file

Note: Please verify skill by going through its instructions before using it.

SKILL.md

name ref-mcp
description Ref.tools MCP server for SOTA documentation search (60-95% fewer tokens than alternatives).
allowed-tools mcp__ref__*
token-budget 500

ref-mcp

Overview

Ref.tools MCP Server

State-of-the-art documentation search with 60-95% fewer tokens than context7/fetch alternatives.

Key benefits:

  • Context-aware deduplication (doesn't repeat docs in same session)
  • Focused snippets instead of full pages
  • 8,500+ libraries indexed with version-specific docs

When to Use

  • Effect-TS: APIs change frequently, training data outdated
  • Any library where unsure of current API
  • Before writing code with external imports
  • When user asks "how do I use X?"
  • Verifying function signatures and parameters

Trigger Phrase

use ref - how do I create an Effect Layer?

Tools

mcp__ref__search

Search for documentation on any library or topic.

mcp__ref__search({
  query: "Effect Layer composition",
  library: "effect-ts"  // optional: focus on specific library
})

Returns focused documentation snippets with:

  • Code examples
  • API signatures
  • Best practices

Usage Patterns

Before Implementing New Feature

1. Search for library-specific patterns
2. Review returned snippets
3. Implement following documented patterns

Effect-TS (Critical)

The Effect API changes frequently. Training data is outdated.

Always query ref before writing Effect code:

mcp__ref__search({
  query: "Effect.gen usage patterns",
  library: "effect-ts"
})

Token Efficiency

Approach Tokens Notes
WebFetch full page 5-20k Includes nav, footer, unrelated content
context7 2-5k Better but still verbose
ref 0.5-2k Focused snippets, session deduplication

Ref automatically deduplicates within a session - repeated queries for same docs return minimal tokens.