Web search that comes back as answers with sources
CrawlGenie Search is a web search API for LLM agents. Ask a question; it searches the live web, opens the best pages, keeps only the passages that answer, and cites each one. Your model reads a few hundred tokens instead of ten pages. Research mode goes deeper and can write a short cited report with your own model key.
One budget for all sources
You set how many tokens the answer may cost; CrawlGenie divides it across the sources it opened.
Every passage has its source
Nothing is summarised away. What comes back is quoted text with the URL beside it.
Fresh, from a real browser
Results come from the live web today, not a cached index.
What happens when you use Search?
Ask
In ChatGPT, Claude, Cursor, or by API: "What do the three main vendors charge per seat?"
Search and open
DuckDuckGo, Brave or Bing results; the best pages open in a real browser; documents are parsed directly.
Keep what answers
Passages are ranked against your question without a model; you get them with sources and a receipt.
What do people use Search for?
Sales and competitive research
Pricing, features, positioning across competitors, with the source for every claim.
Support and success
Answer tickets from vendor docs and status pages without pasting pages into the chat.
Analysts
Compare filings, reports and news for one question; deep mode reads eight sources.
What you get, in detail
Search engines return links; agents need answers. A conventional web search API for LLM agents hands back ten URLs and leaves the model to fetch and read each one, which costs ten pages of tokens for one question. CrawlGenie does the reading on the server: it runs the query in a real browser (DuckDuckGo first, then Brave and Bing), opens the best results, parses documents directly, and ranks every passage against the question without a model. What comes back is a short list of quoted passages, each with its URL, under the token budget you set.
Research mode goes deeper: quick reads three sources, standard five, deep eight, and the response separates evidence (what the pages actually say) from an optional short report written by your own model from that evidence only. Because retrieval needs no model, the evidence step is free of model tokens and deterministic, which makes it suitable for monitoring too: a search watch repeats the query on a schedule and reports new pages only. Every response includes how many tokens were returned and how many the full pages would have cost.
For developers
GET /v1/search?q=…&content=true for results with content; POST /v1/research for cited evidence and an optional report. API reference.
What would this save you?
Move the sliders. The saving is computed from the measured ratios above and the model price you pick.
Assumes 10% of the page is kept per read (measured range 5–15%). Receipts in the product show the real number for your traffic.
Questions people ask
Which search engine is used?
DuckDuckGo first, then Brave and Bing as fallbacks, all through a real browser so nothing needs an API key.
Does Search use a model?
Not to find or rank passages. A model is used only if you ask for a written report, with your own key.
Spend tokens on answers, not on pages.
Create a workspace in a minute. Connect ChatGPT, Claude, Cursor or your own agents. Every answer comes with a token receipt.