mcp-research

A standalone MCP (Model Context Protocol) server providing web research tools. Three battle-tested tools for AI assistants: search the web, fetch & convert pages to markdown, and run compound multi-source research — all via the MCP stdio protocol.

Tools

All tools are read-only — they fetch and transform public web content, never modify anything.

Install

pip install mcp-research

Or run directly with uvx (zero-install):

uvx mcp-research

Configuration

All configuration is via environment variables — no config files needed.

Usage with Claude Code

Add to your Claude Code MCP config (~/.claude/settings.json or project .mcp.json):

{
  "mcpServers": {
    "research": {
      "command": "uvx",
      "args": ["mcp-research"],
      "env": {
        "BRAVE_API_KEY": "BSA...",
        "OLLAMA_URL": "http://localhost:11434"
      }
    }
  }
}

Usage with Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "research": {
      "command": "uvx",
      "args": ["mcp-research"],
      "env": {
        "BRAVE_API_KEY": "BSA..."
      }
    }
  }
}

Tool Details

web_search

web_search(query, max_results=5, summarize=False, auto_fetch_top=False)

Searches the web using a 3-tier cascade for maximum reliability:

1. Brave Search API — fast, high quality (requires BRAVE_API_KEY)
2. DuckDuckGo library — no API key needed, retries on rate limit
3. DuckDuckGo HTML scraper — last-resort fallback

Options:

• summarize: Use Ollama to summarize results (requires running Ollama)
• auto_fetch_top: Also fetch and return the full content of the top result

fetch_url

fetch_url(url, summarize=False, max_chars=50000)

Fetches a URL and converts it to clean markdown:

• SSRF protection: Blocks localhost, private IPs, non-HTTP schemes
• Smart retry: Exponential backoff on 429/5xx, per-hop redirect validation
• 24h cache: SHA-256 keyed, configurable TTL
• Content support: HTML → markdown, JSON → code block, binary → rejected
• Smart truncation: Breaks at heading/paragraph boundaries, not mid-text

research

research(query, depth="standard", context="")

Compound research pipeline:

1. Query rewrite — Ollama optimizes your question into search keywords
2. Web search — finds relevant pages (with zero-result retry expansion)
3. Parallel fetch — fetches top N pages concurrently
4. Summarize — Ollama summarizes each page
5. Synthesize — Ollama produces a final cited answer

Depth levels:

All steps gracefully degrade without Ollama — you still get search results and raw page content.

Development

git clone https://github.com/MABAAM/Maibaamcrawler.git
cd Maibaamcrawler
pip install -e .
python -m mcp_research

License

MIT