Brave Search MCP server setup for Claude Code and Cursor

Quick answer: The Brave Search MCP server gives Claude Code and Cursor access to the Brave Search API, which covers web, news, image, and local search. Drop your API key in the env block, restart your editor, and the model can run live searches. The free tier covers 2000 queries per month. Setup takes about 4 minutes. Tested against server version 0.6.2 on April 15, 2026.

Live search turns Claude from "my training data cuts off at some date" into "I can look up anything right now." For coding work, that means current library versions, breaking change notices, and answers to error messages that did not exist 3 months ago.

This guide covers API key setup, both editor configs, prompt patterns that get the best results, rate limit behavior, and how to chain Brave with a fetch server for full page content.

What this server does

The server exposes 4 search tools: brave_web_search, brave_news_search, brave_image_search, and brave_local_search. Each one takes a query string and optional filters like count, freshness, country code, and safe-search setting. The response is a structured list of results with titles, URLs, and snippet text - about 400 to 600 characters of context per result.

A short list of what the server supports:

• Web search with up to 20 results per call
• News search with freshness filters (past day, past week, past month, past year)
• Image search with up to 50 thumbnails per call
• Local business search with ratings, hours, and map URLs
• Country and language filtering via ISO codes
• Safe search levels: off, moderate, strict

The server does not fetch the pages themselves. It returns snippets. If the model needs full page content, pair it with mcp-server-fetch - one search, then a fetch on the best result.

Installing the Brave Search MCP server

The package is @modelcontextprotocol/server-brave-search. Install is the standard npx -y pattern. Package size is about 2.1 MB.

First, get an API key:

1. Open https://brave.com/search/api and sign up for an account.
2. Pick the Data for AI tier (free up to 2000 queries per month).
3. Copy your API key. It starts with BSA.
4. Add a credit card if you expect to go above 2000 queries per month. Paid tiers start at $3 per 1000 queries.

The free tier is generous enough for a solo developer running a few dozen searches a day. If you have the agent running searches in a loop, track usage on the Brave dashboard - overruns are billed at the standard rate, not blocked.

Configuring for Claude Code

Claude Code reads MCP config from ~/.claude/mcp.json or a per-project .mcp.json. Add a brave-search entry with the API key in env:

{
  "mcpServers": {
    "brave-search": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-brave-search"],
      "env": {
        "BRAVE_API_KEY": "BSAyour_key_here"
      }
    }
  }
}

Restart Claude Code. Run /mcp to confirm the server is attached. You should see 4 tools prefixed with brave_.

For team setups, each developer should have their own API key rather than sharing one. Brave rate limits per key, and a shared key means one busy developer can starve the others. A key per person also makes billing attribution cleaner if you scale past the free tier.

Configuring for Cursor

Cursor reads from ~/.cursor/mcp.json on macOS and Linux or %USERPROFILE%\.cursor\mcp.json on Windows. The JSON is the same:

{
  "mcpServers": {
    "brave-search": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-brave-search"],
      "env": {
        "BRAVE_API_KEY": "BSAyour_key_here"
      }
    }
  }
}

Toggle the server on in Cursor settings. First search takes 2 to 3 seconds on cold start, then search latency settles at 300 to 500 ms per call.

Example prompts and workflows

Search becomes useful when it is specific. A few prompts that work well:

• "Search for the latest Next.js 16 migration guide and summarize the breaking changes."
• "Run a news search for openai pricing changes in the past week and tell me if prices went up or down."
• "Find the top 5 Stack Overflow answers about ENOENT mac npx and which one has the most upvotes."
• "Search for pgvector vs pinecone benchmark 2026 and list the ones with real latency numbers."
• "Find local coffee shops within 2 km of lat 37.7749, lng -122.4194 that are open now."

The model handles the query string generation well. You usually do not need to hand-craft the search syntax. If results are weak, ask the model to "try 3 different query phrasings and compare results" - Brave rewards specific queries more than keyword stuffing.

For research workflows, chain Brave with a fetch server. The pattern is: Claude runs a search, picks the best result, then fetches the full page text. That gets you past the 500-character snippet limit. Both servers together cover about 90% of "look this up for me" use cases.

Troubleshooting

401 Unauthorized. The API key is wrong or missing the BSA prefix. Copy it again from the Brave dashboard. Keys are case-sensitive.

429 Too Many Requests. You hit the rate limit. The free tier allows 1 query per second. Back off for 10 seconds and retry, or upgrade the plan. The server does not auto-retry - the model will see the error.

Results look regional and wrong. Brave uses your IP for geolocation by default. Pass a country parameter in the prompt ("search with country=US") or set the BRAVE_SEARCH_COUNTRY env var in the config to pin results to one region.

Empty results for a common query. Brave's index is smaller than Google's. For very niche or very recent topics, results can be sparse. Pair with tavily-mcp or perplexity-mcp as a backup - you can attach all 3 and let the model pick.

Rate limit counter unclear. Brave returns rate limit headers but the server does not surface them. Check usage on the Brave dashboard directly, or wrap the server in a proxy that logs headers.

Alternatives

Several search MCP servers exist in the wild:

• tavily-mcp is designed specifically for LLM research workflows with a search_and_extract tool that combines search plus page fetch
• perplexity-mcp wraps Perplexity's AI answer engine, which is a different shape - you get a synthesized answer instead of raw results
• mcp-server-fetch is not a search server but pairs well with any of these
• Self-hosted search via a SearXNG instance is an option if you already run one

Brave Search is the right default when you want raw results and a clean free tier. Tavily is better for research workflows where you want extracted page content in one call. Perplexity is better when you want a single synthesized answer with citations. The verdict from 3 months of daily use: keep Brave attached as the general-purpose search, and add Tavily if you find yourself always chaining search plus fetch.