Skip to main content
Glama

Brave Jev MCP

English | 简体中文

Web search results often contain junk articles and passages that mention the right keywords without answering the question. This material consumes context and leaves the AI to find answers among unrelated information.

This project removes unhelpful pages and snippets before search results reach the AI, leaving more context for information that can help answer the question.

Built on Brave Search MCP Server, the project uses Brave to search the web and extract content, and Jev to filter it according to the purpose of the search. Retained content keeps its source links and Brave's original ordering so the AI can read, answer, and cite it.

Filtering currently applies only to brave_llm_context. Other search tools remain available, but their results do not pass through this project's filtering layer or call Jev for screening.

Regular web search (brave_web_search) mainly returns links and short summaries that help people decide which pages to open. In contrast, brave_llm_context provides content snippets extracted from web pages for the AI to read directly. In our testing, brave_llm_context delivers significantly higher information density and more material that can directly support an answer. We therefore recommend brave_llm_context even with filtering disabled.

Installation

Prepare the following:

We recommend setting BRAVE_API_KEY (your Brave key) and TYPESAFE_API_KEY (your Jev key) as environment variables and ensuring that your AI client passes them to the MCP process. The server reads these variables automatically. If your client requires explicit forwarding, use its environment variable settings.

Add the following configuration to your AI client. npx downloads and starts the package automatically:

{
  "mcpServers": {
    "brave-jev": {
      "command": "npx",
      "args": [
        "-y",
        "@romantcig/brave-jev-mcp@latest",
        "--transport",
        "stdio",
        "--enabled-tools",
        "brave_llm_context"
      ]
    }
  }
}

The example uses @latest to get the latest version published on npm.

--enabled-tools selects which tools are exposed to the AI. The configuration above exposes only brave_llm_context. To enable more tools, append their names as separate strings after this argument. If no other tool filtering options are set, removing the argument and its tool names exposes all tools.

Other tools include:

  • Web search (brave_web_search): mainly returns page links and short summaries, with significantly less detailed content than brave_llm_context.

  • Local search (brave_local_search): searches for local businesses and places.

  • News search (brave_news_search): searches recent news articles, supports recency filters, and provides breaking news indicators.

  • Place search (brave_place_search): searches for businesses, places, addresses, and streets using place names or coordinates.

  • AI summaries (brave_summarizer): calls the legacy Summarizer API and requires an existing Pro AI plan subscription. Brave has introduced Answers API as its replacement, but this tool has not integrated it.

  • Video search (brave_video_search).

  • Image search (brave_image_search).

On first launch, the server creates ~/.brave-jev/config.json with filtering enabled by default. Existing configuration is preserved. The Brave key is required for search. Without a Jev key, search still works, Jev filtering is skipped, and results that would otherwise be filtered include setup instructions.

The MCP server loads its startup arguments, environment variables, filtering configuration, and program version at startup. Restart it for changes to take effect.

Related MCP server: groundlink-mcp

Usage

Tell the AI what you need to find. For example:

Use Brave MCP's brave_llm_context to compare SQLite and PostgreSQL. I am choosing a database for a single-machine application. Focus on deployment complexity, concurrent write limits, and suitable use cases, and include sources.

When choosing a database for a single-machine application, concurrency limits and deployment methods are useful; a generic passage about why choosing a database matters contributes little. With Jev filtering enabled, the AI sends search keywords and search intent separately to the tool. You only need to describe what you need.

Default search budgets

To prevent excessive output when the AI omits explicit limits, and to keep sources repeating the same information from filling the context, brave_llm_context uses these defaults:

Parameter

Default

Meaning

count

10

Candidate search results considered when selecting context

maximum_number_of_tokens

4000

Approximate total token budget

maximum_number_of_tokens_per_url

800

Token budget per source

maximum_number_of_urls

5

Maximum number of returned sources

These defaults apply in every filtering mode. Adjust the relevant parameters when calling the tool if you need more sources or detail. count is the number of candidate results, not the number of sources ultimately returned.

Search results first pass through local cleanup rules, then go to Jev for screening against the search intent. If the Jev key is missing or the Jev request fails, the MCP server still returns the content after local cleanup, but Jev's semantic screening is not completed.

Further reading

Available Tools

8 tools
brave_llm_contextbrave_llm_contextA

Search the web with Brave and return pre-extracted, relevance-ranked content chunks (prose, tables, code blocks, structured data). Use the returned content directly when it answers the question; fetch the page only when you need the full context or must verify.

For routine calls, omit the four search budget parameters; the server applies the defaults
shown in their descriptions. Increase a budget only when previous results show that more
source coverage or passage detail is needed. `query` supports site:.

Search rules:
    - Unfamiliar subject: first search the subject name plus the core question for an overview,
      then verify details as needed; do not search for a given detail directly — check whether
      it surfaces naturally in the sources.
    - Keep only keywords directly relevant to the current question in `query`; do not pack
      context and qualifiers into one query.
    - Set `freshness` (pd / pw / pm / py) when recency matters; never hardcode a year in `query`.
    - Two consecutive searches with heavily overlapping results mean the direction is exhausted:
      change the keyword strategy.
    - Search serially; go parallel only when confident of each query.

When relaying results, cite the URL returned with each content item.
Source age, when available, is included with that item; it is not a verified publication date.

Inputs: query (required), count, country, goggles, +5 more.

brave_summarizerbrave_summarizerA

Retrieves AI-generated summaries of web search results using Brave's Summarizer API. This tool processes search results to create concise, coherent summaries of information gathered from multiple sources.

When to use:

- When you need a concise overview of complex topics from multiple sources
- For quick fact-checking or getting key points without reading full articles
- When providing users with summarized information that synthesizes various perspectives
- For research tasks requiring distilled information from web searches

Returns a text summary that consolidates information from the search results. Optional features include inline references to source URLs and additional entity information.

Requirements: Must first perform a web search using brave_web_search with summary=true parameter. Requires a Pro AI subscription to access the summarizer functionality.

Inputs: key (required), entity_info, inline_references.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.2/5.0

    Scored across 8 tools

    Disambiguation4/5

    Each tool targets a distinct search modality or output format (web, images, video, news, summaries, LLM context), so broad intent is usually clear. The main overlap is between brave_local_search and brave_place_search, which both return business/POI data; descriptions differentiate use cases but an agent could still misselect.

    Naming Consistency5/5

    All tools use a consistent brave_ prefix and snake_case, with six following a brave_<type>_search pattern. The two exceptions (brave_summarizer, brave_llm_context) are still clearly named and predictable.

    Tool Count5/5

    Eight tools map cleanly to Brave Search API endpoints, with no redundant or filler entries. The count is well-scoped for a search provider.

    Completeness5/5

    The surface covers all major Brave Search API capabilities: web, local, video, image, news, summarizer, LLM context, and place search. No obvious gap in the core search lifecycle.

    Maintenance

    ActivityMaintained
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers