deep-research-mcp-server

# Deep Research MCP

An [MCP](https://modelcontextprotocol.io) server that gives **Claude, Cursor, or any AI agent** real web-research skills — web search, clean-Markdown page reading, and one-call research dossiers — with **no OpenAI / Perplexity / Tavily key required**. Your agent is the LLM, so it does the synthesis; this server does the hard part (finding sources, fetching them, and stripping pages down to clean, LLM-ready Markdown).

## Tools

| Tool | What it does |
|------|--------------|
| `web_search` | Search the web → ranked `{title, url, snippet}`. |
| `read_url` | Fetch a URL → main content as clean Markdown. |
| `deep_research` | Search → read the top N sources in parallel → a citation-ready dossier. |

## Install

### Claude Desktop / Cursor / Windsurf

Add to your MCP config (`claude_desktop_config.json`, Cursor `mcp.json`, etc.):

```json
{
  "mcpServers": {
    "deep-research": {
      "command": "npx",
      "args": ["-y", "deep-research-mcp-server"]
    }
  }
}
```

Then ask your agent: *"Use deep_research to compare the top open-source vector databases in 2026 and summarize with citations."*

## Why this over Tavily / Firecrawl?

- **No API key** — search and page-reading run locally; your agent supplies the intelligence.
- **Agentic** — `deep_research` orchestrates multi-source reading and returns a citation-ready dossier, not a single page.
- **Free & open** — MIT licensed, `npx` and go.

## Notes

- Uses public metasearch (SearXNG) with a DuckDuckGo fallback; very fresh pages may not be indexed yet.
- Some sites block automated fetches; `deep_research` skips failed sources and reports how many it read.

---

Also available as a hosted Apify Actor: [apify.com/renzomacar/deep-research-mcp](https://apify.com/renzomacar/deep-research-mcp). Built by [Renzo Madueño](https://github.com/Perufitlife).