How is this different from Tavily or Exa?

SerpKite returns Google's live results (or five other engines with engine=), reads the top pages in the same call and ranks passages without a model, so results are reproducible. When you want text written for you, cited answers and research reports are built on those same results. Credits never expire and failed searches are free. See SerpKite vs Tavily and SerpKite vs Exa.

Does SerpKite generate answers with an LLM?

Only when you ask. By default it returns sources, page content and ranked passages, and your model writes the answer. With include_answer, /v1/answer or /v1/research, Claude Haiku 5.5 writes it from the results, and every claim cites a numbered source. A failed answer is not billed.

Which agent frameworks are supported?

A remote MCP server for Claude, Cursor, ChatGPT and other MCP clients, official TypeScript, Python and Go SDKs, and guides for LangChain, LangGraph, CrewAI, LlamaIndex, the Vercel AI SDK, the OpenAI Agents SDK, Pydantic AI, Mastra, n8n, Make, Zapier, Dify and Flowise.

How many tokens does a search use?

Depends on the format and options. Check the X-Tokens-Estimate header, or compare JSON, compact and Markdown for your own query with the SERP token counter.

Agents · RAG · MCP

Web search API for AI agents

An agent needs more than ten blue links: the pages behind them, in a form that fits a context window, from a tool that doesn't fail when one search engine does. SerpKite does all of that in one call.

Get a free API key Tool calling for agents guide

How it works

Search, read and cite: everything an agent's web tool needs.

A web tool for an agent has three jobs: find sources, read them, and hand the model something small enough to reason over. /v1/search does the first with Google's live results. include_content: 1…5 does the second in the same call, fetching the top result pages as Markdown. highlights: true does the third, keeping only the 3 passages per page that match the query, ranked with BM25.

The rest keeps it running in production. format: "markdown" or "compact" and fields cut the tokens, and X-Tokens-Estimate reports the size before you prompt. engine: "auto" falls back to Brave, Bing, DuckDuckGo, Yahoo or Yandex when Google is blocked, and meta.engine names the source. Failed, blocked and empty searches are free.

Use it as a remote MCP server, through the TypeScript, Python and Go SDKs, or as a native tool in LangChain, CrewAI, the Vercel AI SDK and the OpenAI Agents SDK.

One tool call: search, read, cite

cURL

TypeScript

Python

curl https://api.serpkite.com/v1/search \
  -H "Authorization: Bearer $SERPKITE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"q":"how do vector databases handle deletes","include_content":3,"highlights":true,"engine":"auto"}'

What the model receives · illustrative

{
  "results": [
    {
      "position": 1,
      "title": "Deletes and compaction",
      "link": "https://docs.example.org/concepts/deletes",
      "domain": "docs.example.org",
      "published_at": "2026-09-12",
      "highlights": [
        {
          "text": "Deletes are written as tombstones and removed from the index during compaction…",
          "score": 8.104,
          "heading": "Deletes and compaction > Tombstones"
        }
      ]
    }
  ],
  "meta": {
    "engine": "google",
    "credits_used": 4,
    "cached": false
  }
}

Each agent job, one parameter

Each agent job, one parameter
The agent needs to…	SerpKite	Cost
Find sources	`/v1/search`, news, scholar, maps…	1 credit per page
Read the top results	`include_content: 1…5`	+1 credit per page read
Keep only what answers the question	`highlights: true` (BM25, no model)	Free
Fit the context window	`format: "markdown"`, `fields`, `X-Tokens-Estimate`	Free
Keep working when Google blocks	`engine: "auto"`	Failed attempts free
Cross-check sources	`engine: "consensus"`	1 per engine that answered
Read a whole site or a URL list	crawl, extract, map	1 credit per page
Get a cited answer	`include_answer: true` or `/v1/answer`	+2 credits (+4 over fetched pages); free if the model fails
Turn pages into JSON	`output_schema` on extract	+4 credits per page
Research a question end to end	`/v1/research`	30–200 credits per report

Why it matters