Brave Search MCP Server

Overview

Brave Search MCP Server wraps the Brave Search API as a set of Model Context Protocol tools, giving agents privacy-respecting access to web search, local business/place search, and AI-powered summarization without operating a custom crawler. It lives in the official `modelcontextprotocol/servers` monorepo (brave-search subdirectory) and is one of the most widely deployed reference MCP servers. Agents call the `brave_web_search` and `brave_local_search` tools to retrieve structured results they can reason over, cite, or pass to downstream tools. It works best with Claude Desktop and any MCP-compatible agent host, including Cursor, Windsurf, and custom agent runtimes that speak the MCP stdio transport.

Best for

Best for
Claude Desktop users and MCP-compatible agents that need privacy-respecting web and local search without managing a crawler

Use cases

• Give a Claude Desktop agent real-time web search without a custom backend
• Run local business or place lookups inside an agent workflow
• Bootstrap a research agent that needs general web context before reasoning
• Combine with a fetch/browser MCP for full search-then-read loops

Notes

What it does

Brave Search MCP Server translates the Brave Search API into MCP-native tools that any compatible agent can call directly. The two primary tools are brave_web_search (general web results with optional freshness and country filters) and brave_local_search (business and place results with addresses, ratings, and hours). An optional summarization surface lets agents request an AI-generated answer alongside raw results.

How to deploy

Install via npx or add it to your Claude Desktop claude_desktop_config.json under mcpServers. Set the BRAVE_API_KEY environment variable and the server is ready. No Docker, no persistent process — it runs as a child process spawned by the MCP host on demand.

Best practices

Pair with a fetch or browser MCP (Playwright, Puppeteer) for a full search-and-read loop: search first to find the right URLs, then fetch the full page content for deep reasoning. Cache repeated queries to stay within free-tier rate limits.