DeepSift Search: Tavily-compatible web search for AI agents

What does DeepSift Search do?

DeepSift Search gives your AI agent live web search. You send one question; it searches Google, opens the best pages, and returns only the passages that answer the question, ready to drop into an LLM prompt.

It speaks the same request and response format as the Tavily API. If your code already uses Tavily, you can switch by changing one line: the base URL.

Use it three ways:

• 🔁 As a Tavily drop-in: point the Tavily SDK (or any HTTP client) at DeepSift's address
• 🤖 As an MCP server: add it to Claude, Cursor, Windsurf or any MCP client, and get web_search, web_read, web_map and web_crawl tools
• 📋 As a normal Apify run: paste a list of search queries, get a table of results

Why use DeepSift Search?

• Fewer tokens for your LLM. You get the best few passages from each page, not whole pages of menus and cookie banners.
• Switch in a minute. Same paths (/search, /extract, /crawl, /map), same fields, same response shape as Tavily.
• Pay per use, no subscription. You pay per search on your Apify account. Failed searches are free.
• Real Google results. Up-to-date pages, with country, language, date-range and website filters.
• No extra API keys. Your Apify token is the only key you need.

How to use it as a Tavily drop-in

Your address is https://aadstra--deepsift-search.apify.actor. Use your Apify API token as the API key.

Python (Tavily SDK)

from tavily import TavilyClient

client = TavilyClient(
    api_key="<YOUR_APIFY_TOKEN>",
    api_base_url="https://aadstra--deepsift-search.apify.actor",
)
result = client.search("who acquired Tavily in 2026", max_results=5)
for row in result["results"]:
    print(row["url"], row["content"])

Any language (HTTP)

curl -X POST https://aadstra--deepsift-search.apify.actor/search \
  -H "Authorization: Bearer <YOUR_APIFY_TOKEN>" \
  -H "Content-Type: application/json" \
  -d '{"query": "best open source vector database 2026", "max_results": 5}'

Supported endpoints:

Endpoint	What it does
POST /search	Search Google; returns the best passages from each result page
POST /extract	Readable text of up to 20 pages or PDFs (optionally only the parts about a query)
POST /crawl	Reads many pages of one website; instructions keeps the relevant ones
POST /map	Lists a website's page addresses

How to use it as an MCP server

Add this to your MCP client's settings (Claude Code, Claude Desktop, Cursor and others):

{
  "mcpServers": {
    "deepsift-search": {
      "url": "https://aadstra--deepsift-search.apify.actor/mcp",
      "headers": { "Authorization": "Bearer <YOUR_APIFY_TOKEN>" }
    }
  }
}

Your agent then has four tools: web_search, web_read, web_map and web_crawl.

How to use it as a normal run

1. Click Try for free.
2. Type one search per line in Search queries.
3. Pick how many results you want and, if you like, a time range, country or website filter.
4. Click Start. Results appear in the Output tab and can be downloaded as JSON, CSV or Excel.

Input

Field	Meaning
queries	List of searches (required)
searchDepth	basic (best passage per page), advanced (up to three passages, longer page reads) or fast (Google snippets only)
maxResults	Results per search, 1 to 20 (default 5)
topic	general, news or finance
timeRange	day, week, month or year
includeDomains / excludeDomains	Only search, or never return, these websites
country / language	Two-letter codes, e.g. us, in, de / en, hi
includeFullPages	Also return each page's full readable text

Output

One row per result:

{
  "query": "best open source vector database 2026",
  "rank": 1,
  "title": "The top open source vector databases & comparison",
  "url": "https://redis.io/blog/best-open-source-vector-databases-comparison/",
  "content": "The top open source vector databases & comparison · Redis · Milvus · Weaviate · Qdrant ...",
  "score": 0.93,
  "publishedDate": null
}

How much does it cost?

You only pay for what you use. Prices per event:

Event	Price
Search (basic or fast)	$0.004 ($4 per 1,000)
Advanced search	$0.008 ($8 per 1,000)
Page read (each page returned by /extract or /crawl)	$0.001 ($1 per 1,000)
Site map (each /map request)	$0.002 ($2 per 1,000)

Platform usage is included: you pay only the prices above. Failed requests are not charged. You can set a maximum cost for each run, and DeepSift stops (and stops charging) when it is reached.

For comparison, Tavily's published pay-as-you-go price is $0.008 per credit, and an advanced search uses two credits.

Limits and good to know

• No written answers or research reports. include_answer returns null and /research is not available. DeepSift returns the sources; your own LLM writes the answer.
• Speed. The first request after a quiet period takes about 5 to 10 seconds while the server starts. After that, a search takes a few seconds (it opens and reads every result page). Repeating the same search within 24 hours is answered from a cache.
• Pages that only show text after JavaScript runs may come back thin.
• Up to 20 results per search. More than 10 reads a second Google results page.

FAQ

Is it legal to use? DeepSift reads public web pages and Google results pages. You are responsible for how you use the data. Don't use it to collect personal data without a legal basis (for example under GDPR).

Is DeepSift open source? Yes. The self-hosted version for your own computer is free on GitHub. This Apify version saves you from running and maintaining it yourself.

I found a bug or need a feature. Open an issue in the Issues tab. We usually reply within a few days.