Brave's search API, now with more MCP bureaucracy

What it does

This is Brave’s official MCP server — a TypeScript adapter that exposes the Brave Search API as a set of structured tools for LLM agents. It covers web, image, video, news, and local search, plus a dedicated brave_llm_context endpoint that returns pre-extracted content optimized for RAG pipelines. There’s also a two-step summarizer: first search with summary: true, then feed the returned key into brave_summarizer for an AI-generated digest.

The interesting bit

The 2.x release deliberately broke backward compatibility by switching the default transport from HTTP to STDIO, following “established MCP conventions.” They also stripped base64-encoded image data from responses after realizing it was burning context windows for no good reason. It’s rare to see an official vendor adapter admit its own bloat and actually cut it.

Key highlights

• Eight search tools, from vanilla web search to geofenced place lookup with opening hours and ratings
• brave_llm_context with fine-grained token and snippet budgets (up to 32K tokens, 256 snippets)
• Tool-level allowlists and blocklists via BRAVE_MCP_ENABLED_TOOLS / BRAVE_MCP_DISABLED_TOOLS
• Stateless HTTP mode for Amazon Bedrock Agentcore
• Pro plan gating on local search and extra snippets

Caveats

• Requires a Brave Search API key; some features (local search, extra snippets) are Pro-only
• The README is thorough on parameters but vague on error handling and rate limits
• Image search no longer returns displayable data — just metadata and URLs

Verdict

Worth a look if you’re already committed to MCP and want a privacy-tilted search backend with decent RAG support. Skip it if you just need a quick search wrapper; a direct API call is less indirection for simpler use cases.