# SERPdive MCP Server

Give any MCP client real-time web search with answer-ready results. [SERPdive](https://serpdive.com) is the AI Search API: ask a question, get the actual content of the best pages, extracted, cleaned, and sized for an LLM. On a [public, replayable 1,000-question benchmark](https://github.com/edendalexis/serpdive-benchmark), SERPdive runs at the same speed as Tavily, feeds your LLM 20.2% fewer tokens, and wins 60.7% of decided quality duels. If you are evaluating Tavily alternatives, that benchmark is public and replayable end to end: same questions, same judge, your machine.

**There is a free tier, and it has no ceiling.** The `krill` model is free and unlimited under fair use — no card, no credits, nothing to decrement. It returns the shortest set of sentences that still answers (about 700 tokens a search, roughly half what the usual alternatives send), one request at a time, at low priority. Use it to build; switch one word to `mako` when you need depth and steady latency.

One tool, `serpdive_search`. Two ways to run it:

- **Hosted (recommended)**: `https://mcp.serpdive.com`, nothing to install.
- **Local (this package)**: `npx -y serpdive-mcp`, stdio, zero dependencies.

Get a free API key at [serpdive.com/dashboard/keys](https://serpdive.com/dashboard/keys) (no card required).

## Hosted server

### Claude Code

```bash
claude mcp add --transport http serpdive https://mcp.serpdive.com \
  --header "Authorization: Bearer sd_live_YOUR_KEY"
```

### Cursor and other JSON-config clients

```json
{
  "mcpServers": {
    "serpdive": {
      "url": "https://mcp.serpdive.com/?key=sd_live_YOUR_KEY"
    }
  }
}
```

Both `https://mcp.serpdive.com/` and `https://mcp.serpdive.com/mcp` answer, so either URL shape works.

## Local server (npx)

### Claude Desktop

Add to `claude_desktop_config.json` (Settings > Developer > Edit Config):

```json
{
  "mcpServers": {
    "serpdive": {
      "command": "npx",
      "args": ["-y", "serpdive-mcp"],
      "env": { "SERPDIVE_API_KEY": "sd_live_YOUR_KEY" }
    }
  }
}
```

### Claude Code

```bash
claude mcp add serpdive --env SERPDIVE_API_KEY=sd_live_YOUR_KEY -- npx -y serpdive-mcp
```

### Cursor and other JSON-config clients

```json
{
  "mcpServers": {
    "serpdive": {
      "command": "npx",
      "args": ["-y", "serpdive-mcp"],
      "env": { "SERPDIVE_API_KEY": "sd_live_YOUR_KEY" }
    }
  }
}
```

## The tool

`serpdive_search(query, model?, max_results?)`

| Argument | Type | Description |
|---|---|---|
| `query` | string, required | The search, in any language. Localization is automatic. |
| `model` | `"mako"` \| `"moby"` | `mako` (default): the fact-carrying sentences of each page, fast. `moby`: full page text — slower (~1s more) and several times more tokens; use only when mako's key sentences are insufficient (long-document analysis, deep research). |
| `max_results` | integer, 1-10 | Cap on delivered results. Omit for the engine's calibrated mix. |

The response is the raw SERPdive JSON: `query`, `model`, `response_time_ms`, optional `extra_info`, and `results` as `[{ url, title, date?, content }]`. Failed searches are never billed.

There is no `answer` argument here, on purpose. Over MCP the consumer is always an
LLM, which can write its own answer from the extracted content — a second,
server-side synthesis costs a round-trip on every search and buys nothing. The
[API](https://serpdive.com/docs) still offers `answer: true` for callers that are
not models.

## Pricing and limits

A `mako` search costs 1 credit, `moby` 1.5. Every account gets free monthly credits, no card required. Full reference: [serpdive.com/docs](https://serpdive.com/docs).

## License

MIT