Open source web search API in 2026

Keep the open-source client, drop the server

Many open-source AI apps take a SearXNG address for web search. SpecStream answers that call at one address, so the app keeps its SearXNG setting and gets results[] back as before. Where the setting takes the whole search address, use https://api.specstream.ai/searxng/<key>/search. The apps set up this way are on web search for your AI app.

# Before: your own instance
curl "http://localhost:8080/search?q=open+source+vector+databases&format=json"

# After: nothing to host
curl "https://api.specstream.ai/searxng/$SPECSTREAM_API_KEY/search?q=open+source+vector+databases&format=json"

The same key answers the open-source scrapers too: Firecrawl's search, scrape, map and crawl at https://api.specstream.ai/firecrawl, and a Crawl4AI server's /crawl, /crawl/stream and /md at https://api.specstream.ai/crawl4ai. The step by step move from an instance is on Migrate from SearXNG.

No IP of yours for the engines to block

A self-hosted search API asks Google, Bing and DuckDuckGo from your server's address. SearXNG suspends an engine for a day after a CAPTCHA, OpenSERP's README says browser mode does not stop CAPTCHAs, and DonSeTch's says keyless search rate-limits without a proxy. With SpecStream the engines never see your address, and a call stopped by a busy search costs nothing.

The most accurate web search API

SpecStream scores 97.9 F-score on all 1,000 official SimpleQA Verified questions (978 correct), and 64.0 item F1 and 43.4 row F1 on all 200 official WideSearch tasks. The Extract API scores 0.908 F1 on WCXB. Each score and its method is on the benchmarks page.

The privacy you self-hosted for, set on the key

Zero retention is a field on each key: a search made with it keeps no request text and no answer after the reply, and the charge keeps only the account, the key, the request ID, the price and the time. A second field, eu_egress_only, sends every search from our server in Madrid, straight to each engine, never through a proxy in another country.

POST https://api.specstream.ai/v1/account/keys

{
  "name": "search",
  "scopes": ["account:read", "finds:spend"],
  "spend_cap_cents": 500000,
  "zero_retention": true,
  "eu_egress_only": true
}

Search, pages and rows on one key

The Extract API returns the main text of up to 20 pages or documents (PDF, Word, Excel) a call at $0.80 per 1,000 pages, and a page with no text is not charged. Deep Search returns every match on the live web as structured rows at 5¢, 15¢ or $1 a row, with the most a run can cost quoted before it starts. Agents connect to https://api.specstream.ai/mcp.

SpecStream

The live web data engine. Its Search API answers SearXNG's JSON at https://api.specstream.ai/searxng/<key> for $1 per 1,000 searches, with nothing to host, and scores 97.9 F-score on all 1,000 SimpleQA Verified questions. The Extract API reads pages at $0.80 per 1,000, and Deep Search returns structured rows at 5¢, 15¢ or $1 a row.

SearXNG

An AGPL-3.0 metasearch engine you host with Docker or Podman and a Valkey database. JSON is off until the admin adds it to the formats list, and a CAPTCHA suspends an engine for 86,400 seconds by default. SpecStream answers the same format=json call on every key.

OpenSERP

An MIT REST API and CLI over six engines, in one Go binary or Docker image, that loads each results page in Chromium and reads up to 20 URLs in a batch. You cover compute and proxies, or pay $0.0078 a credit on its Cloud. SpecStream searches and reads pages on one key, with no browser of yours to run.

DDGS

A Python metasearch library with ten text backends, from Bing and Brave to Mojeek and Yahoo, and an API server on port 4479. Its README says it is for educational purposes only. SpecStream is a paid API with a key, a spending limit and a request log for every call.

DonSeTch

An AGPL-3.0 Rust binary that fetches, searches, crawls and takes screenshots with no keys, across 10+ search backends. Its README says keyless search rate-limits without a proxy and that CAPTCHA walls need an unlocker key. SpecStream reads pages through the Extract API, retrying a refused page as a browser, then from the site's feed, then from its newest archive capture.