Rather than a tool per search engine there is a single call whose parameters change with the engine, and the per-engine schemas are published as resources the model can read rather than guess at. What you spend on context is a choice at call time: compact JSON drops metadata, markdown cuts the token count sharply, and on hosts that support interactive UI the opt-in tools hand the results to the host to render instead of into the conversation.
SerpApi's hosted search server. A single search tool takes the parameters of any SerpApi engine, so a live query on Google, Bing, Yahoo, DuckDuckGo, YouTube, eBay, Amazon, Google Scholar or the App Store comes back as structured results.
- One search tool covering every SerpApi engine and result type, with the parameters changing per engine
- Per-engine parameter schemas published as resources at serpapi://engines and serpapi://engines/<engine>, so the model can look up what an engine accepts
- Answer boxes, organic results, news, images and shopping detected and formatted automatically — weather and stock questions answered through the same call
- JSON in complete or compact mode, where compact strips the metadata
- Markdown output, which the vendor measures at about half the tokens on average and more than 90% fewer for engines with deeply nested JSON
- Two opt-in tools, search_table and search_dashboard, that render results as an interactive table or dashboard in hosts supporting the MCP Apps extension — so the bulk SERP JSON never enters the model's context
- A self-hostable build for running the same server yourself
A SerpApi API key. The hosted server is at https://mcp.serpapi.com and takes the key either in the URL path or as an Authorization: Bearer header — path-based is the vendor's recommendation. Self-hosting needs Python 3.13 or newer and uv, or the Docker image, and then answers on your own host. MIT.
One command plus a key — claude mcp add --transport http serpapi https://mcp.serpapi.com/YOUR_SERPAPI_API_KEY/mcp, then supply credentials