An agent can write a convincing answer about a product that changed its pricing yesterday. Convincing isn’t current. Search can point the agent toward a price page, but the answer still depends on which page it opened, when that page was updated, and whether the price actually applies to the user asking.
Brave Search MCP is one way to put search inside that workflow. Brave maintains an official server repository that connects the Brave Search API to MCP clients. This spotlight looks at the server’s actual interfaces, not a promise that search will make every answer true.
What the server provides
The official server exposes separate tools for web, news, image, video, local and place search, along with an LLM-context tool and a summarizer. You can choose the tool that matches the question rather than ask a single generic search interface to do everything. A coding assistant checking a changed library API needs the reference page; a research agent tracking a launch may start with news results filtered by recency. The summarizer isn’t a substitute for either source.
For example, brave_web_search accepts a query, country and language preferences, a result count, and a freshness value such as pd for the past day or pw for the past week. These filters constrain the search request; they do not prove that a returned page was published in that period or that its contents are accurate. The agent should inspect the result page and its publication metadata before citing it.
brave_llm_context retrieves pre-extracted web material with token and URL limits. It is useful for finding candidate passages, but extraction can strip away the part of a page that changes their meaning: a pricing footnote, a version selector, a correction. Open the original URL before relying on a passage for a consequential decision.
Where it fits in an agent workflow
Imagine a developer agent asked whether a payment SDK still requires a particular initialization option. A safe sequence is:
- Search for the SDK’s current official documentation and release notes.
- Prefer the vendor’s versioned reference over third-party tutorials.
- Open the source page and confirm the option against the version in the project’s lockfile.
- Report the source URL and version, and mark any remaining uncertainty.
The lockfile check is the important part. Search can’t do it for you, and finding a plausible documentation page is not a test of the code in your project. An MCP tool makes the query available to the agent; the host and application still decide which tools and credentials it can use. For the broader architecture, see our MCP tools explainer and guide to evaluating MCP servers before installation.
Already have the documentation URL? Fetch it directly. Use search when you don’t know where the answer lives; use a browser or crawler instead when you need structured extraction across pages you already know.
Installing the current official package
The official README documents the @brave/brave-search-mcp-server package. In its current 2.x instructions, stdio is the default transport. A local MCP client configuration can look like this:
{
"mcpServers": {
"brave-search": {
"command": "npx",
"args": ["-y", "@brave/brave-search-mcp-server", "--transport", "stdio"],
"env": {
"BRAVE_API_KEY": "REPLACE_WITH_YOUR_OWN_KEY"
}
}
}
}Create a Brave Search API account and key first; check your client’s configuration format before copying this example. Keep the actual key out of repositories and shared screenshots. The AgentNDX directory entry currently lists an older @modelcontextprotocol/server-brave-search command, which also appears in an older Brave Claude Desktop guide. For a new installation, follow the current official server README rather than assuming both packages have the same tools or release behavior. We have not run an authenticated search or benchmarked the server here.
Costs and security boundaries
A Brave API key is required. The server’s documented capabilities are not a guarantee that every tool works under every subscription: the README says full local search capabilities need a Pro plan and extra snippets are Pro-only. Check the current plan and billing terms in your own account before building a workflow around those features. A high-frequency background agent can consume API allowance even when no human is reading its results.
The server also supports an HTTP transport, but its README warns that the HTTP endpoint is unauthenticated and binds to loopback by default. Do not expose it to the public internet simply to make it reachable from another machine. Use an appropriate access boundary and review the server’s host and origin settings if remote access is genuinely required.
Treat result snippets and extracted page text as untrusted data. A web page that tells the agent to ignore its task or reveal a key has not acquired authority by appearing in a search result. Separate source material from instructions, and keep external sends or writes behind explicit approval. Our article on MCP tool results as data, not instructions goes deeper on that boundary.
The decision
Use Brave Search MCP when the agent needs to locate public sources inside an MCP workflow and region, language or recency controls will help narrow the search. If the sources are already known, skip the extra search call. Otherwise, choose only the tools the job needs, put a budget on queries and retain the URLs you actually used.
FAQ
Does Brave Search MCP browse and verify every returned page?
No. Search tools return results or extracted context, depending on the tool. Your workflow must still open and assess the sources behind claims that matter.
Can I use the server without a Brave Search API key?
The official server requires BRAVE_API_KEY or a key file. Check Brave’s current plan terms before relying on a particular allowance or feature.
Should I use stdio or HTTP?
For a local client launching the server as a subprocess, use the documented stdio default. HTTP is an option for a different deployment topology, but it requires careful network and access controls.