Live web, news, image, video, and local search through Brave Search with agent-friendly result retrieval.
Looking for more MCP servers? Browse the full MCP tools directory or explore more tools in Search.
Brave Search MCP is the precision search primitive every agent research stack needs. Its strength is exactness and vertical control; its weakness is that search results and snippets are not evidence. Pair it with Fetch or Firecrawl, enforce domain and freshness policies, and require source-passage verification before publishing. It complements Exa rather than replacing it.
Brave Search MCP is the literal-retrieval counterweight to neural search: it is the tool an agent needs when exact strings matter more than conceptual similarity. Package names, CVE identifiers, error messages, API paths, release tags, domain-restricted documentation, and proper nouns all punish semantic paraphrase; a neural engine may return something related while missing the exact token that solves the incident. Brave exposes search primitives across web, news, images, video, and local results, allowing the agent to choose the index rather than pretending every query is a webpage question. The architectural nuance is result-to-evidence separation: search snippets are ranking artifacts, not authoritative content. A robust agent uses Brave to discover candidates, fetches the primary page, verifies the exact passage and date, and records the final URL. The silent failure that dominates search automation is query-type mismatch — asking a broad web index for a local business, or a news index for an API symbol, produces valid-looking irrelevant results. Add domain filters, freshness windows, result caps, and a fetch step. Brave is excellent at finding the needle; it is not proof that the needle says what the model claims.
The teams and workflows that benefit most from this tool.
Exact error and package lookup
Current documentation and release research
News, image, video, and local search
Domain-restricted SEO and competitor research
Choose Brave Search MCP Server when fresh, real-time information matters and your AI assistant needs live search or research capabilities.
Good fit
Things to watch before choosing this tool.
Snippets can misrepresent page meaning
Requires a second retrieval step for evidence
Keyword retrieval is weaker for abstract conceptual discovery
Unbounded searches can consume API quota
Use the following configuration as a starting point for Claude Desktop or any compatible MCP client, then replace placeholder credentials with your own values.
{
"mcpServers": {
"brave-search": {
"command": "npx",
"args": ["-y", "@brave/brave-search-mcp-server"],
"env": {
"BRAVE_API_KEY": "BSA_your_api_key"
}
}
}
}These setup steps cover the typical installation flow for this MCP server.
Create a Brave Search API key and store it as BRAVE_API_KEY; never expose it in prompts or committed configuration.
Configure the server and choose the correct search vertical — web, news, images, video, or local — instead of sending every query to the generic web index.
Pro-Tip: The fatal flaw 90% of developers commit is publishing from search snippets. Snippets are truncated, query-dependent, and may combine text from different page sections; a valid URL does not validate the claim. Use Brave for discovery, then fetch the primary source, verify the exact passage and publication date, and cite the page rather than the snippet.
Use includeDomains and excludeDomains for technical research, and set freshness filters for releases, pricing, vulnerabilities, and news.
Cap result counts and cache repeated queries; unrestricted agent loops can spend quota while repeatedly rediscovering the same sources.
Verify with an exact error string, a domain-restricted documentation query, and a freshness-sensitive news query.
Supported environments for Brave Search MCP Server.
Claude Desktop
Yes
Cursor
Yes
VS Code
Yes
API-key-based search server. Use the correct vertical, domain filters, freshness controls, and fetch verification before citing results.
Common questions for Brave Search MCP Server.
The silent failure is snippet overreach. Brave returned a valid result and a short excerpt, but the model expanded that excerpt into a broader claim or merged it with another result. The engineering fix is to treat snippets as discovery only: fetch the page, quote the supporting passage, verify the date and domain, and reject claims that cannot be located in the source.
Brave is stronger for exact strings, navigational queries, current news, and domain filtering; Exa is stronger for conceptual neural retrieval and semantically similar sources. A serious research workflow uses Brave for literal lookup and Exa for corpus expansion.
Local search can discover businesses and places, but detailed routing and geocoding require a maps-specific integration. Do not treat a local result as a complete navigation answer.
They constrain retrieval toward recent content, but they do not guarantee that the page itself is current or authoritative. Verify the publication date and update timestamp on the source page.
| Feature | Brave Search MCP Server | Competitor |
|---|---|---|
| Exact-match retrieval | ✅ Strong for identifiers, error strings, and package names | ❌ Neural retrieval may return related but non-identical concepts |
| Search verticals | ✅ Web, news, images, video, and local indexes | ❌ Single general-purpose web index |
| Evidence quality | ❌ Requires a fetch-and-verify step after discovery | ✅ Content-inline retrieval can reduce the second hop |
| Conceptual similarity | ❌ Keyword-oriented retrieval | ✅ Embedding-based semantic search |