ref-tools-mcp
Basic Information
Ref MCP is a ModelContextProtocol (MCP) server designed to give AI coding tools and agents access to accurate, up-to-date documentation for APIs, services and libraries in a token-efficient way. It is intended to be a single place to surface the exact documentation context an agent needs to answer coding and integration questions while avoiding large noisy document dumps. The server matches how models search, tracks session search trajectories, filters repeated results, and returns focused excerpts of documentation rather than entire pages. The repository provides both a legacy stdio server and instructions for a recommended streamable HTTP setup so MCP clients can call Ref as a tool in development or production agent workflows.
Links
Stars
288
Language
Github Repository
Categorization
App Details
Features
Agentic search that mirrors model search behavior and minimizes context by selecting exactly relevant resources. Session tracking and filtering to avoid returning duplicate or redundant results across repeated searches. Partial-page fetching that extracts the most relevant ~5k tokens of a document to reduce noise and context rot. Two primary tools: ref_search_documentation for querying public and private documentation including web, github, repos and PDFs, and ref_read_url to fetch a URL and convert it into markdown with focused content. OpenAI-compatible tool naming for deep research workflows. Development tooling includes npm scripts, a dev watcher, a build step, and optional MCP Inspector integration for debugging.
Use Cases
The project reduces token usage and cost by returning only the documentation fragments an agent needs, which also helps avoid model degradation caused by excessive context. It speeds up agent workflows by performing focused searches and returning concise, relevant snippets or code examples rather than whole large pages. It supports both streamable HTTP and stdio deployments so developers can integrate Ref into local and hosted agent stacks. The tools help coding agents find factual answers and code snippets, support deep research scenarios, and provide OpenAI-compatible tool mappings for clients that require specific tool definitions.
-
$$$$ -
$$$$