Independent web search
The open web, indexed by us.
An independent search engine: an index we keep ourselves, a crawler of our own, and an answer an agent can read — with no ranking sold to anyone. It works in a browser at search.snaga.ai and over HTTP, MCP and the addresses Ollama's clients already speak.
Testinga public beta at search.snaga.ai — thirty searches a minute without a key
How to connectDetective
Finds the details that help you decide.
Our own index
About 397 million pages in more than three hundred languages, in an index we keep ourselves and grow with our own crawler (measured 2026-10-22).
What the page says about itself
Every result carries the page's title and description, its site, its language and the day we crawled it — and no text beyond what the page allows. An answer also names the query words a result has and the ones it lacks; a query of operators alone says no_words, so an agent can tell a miss from a question with no words in it.
Robots and opt-outs
robots.txt as RFC 9309 defines it, noindex, nofollow, nosnippet and crawl-delay are kept. SnagaBot names itself in every request, and a site or a page can be taken out of the index by asking.
How to connect
In a browser
search.snaga.ai — thirty searches a minute per address, ten results, no key, and the same search works without JavaScript.
Open the engineOver HTTP
GET /v1/search with q, limit (1–50), lang and site. A key comes by mail: 5,000 requests a month free, then $1 per 1,000. Errors are never billed.
API for agentsThrough MCP
https://search.snaga.ai/mcp with the key as a bearer token; one tool call is one search, and the tool's answer is text with the results beside it.
Ollama's addresses
POST /api/web_search and POST /api/web_fetch take the arguments and return the answers an Ollama client already expects: only the address changes.
Before you point anything at it
It is a beta, and the index is a snapshot that grows with each crawl: the date on a result is the day we crawled it, not the day it was published, and News means pages crawled in the last fourteen days. Results are not filtered by topic. Queries are never recorded; a usage record holds the time, the key, the address that was called, the number of results, the duration and the price.