What is this?

4got is a marketplace for questions and hypertext answers. It fans out your query to hundreds of search engines simultaneously, merges the results, and strips away all the tracking. Over 50 oracle types give instant answers for weather, stocks, crypto, dictionary, recipes, flights, and more. An AI research agent can synthesize deeper answers from real sources. No ads. No accounts. No JavaScript required for core functionality.

(under construction)

Why

(under construction)

Where this is going

(under construction)

How it works

Your search goes to every enabled engine at once. Results come back, get deduplicated and scored by cross-engine consensus — results that multiple engines agree on float to the top. Engine weights, position decay, and intent-based boosts (code, academic, images) shape the final ranking. The engine timeline shows exactly which engines responded and how long they took.

Oracles check if your query matches a known pattern — a calculation, a unit conversion, a stock ticker, a weather lookup — and return an instant answer right in the result page. If you search the same query twice, the AI research agent activates and synthesizes a comprehensive answer by fetching and reading the top result URLs.

Friend instances share cached results over a federation protocol. Before fanning out to upstream engines, 4got checks local cache, then friend caches. This minimizes upstream load across the network and means popular queries are often answered without touching an external engine at all.

What leaves this server

A search engine is a kind of AI, and so are many of the instant answers. In time we expect much of this to run on your own machine, with nothing sent to anyone's server. The web itself is different: for the foreseeable future there will be no good search engine for the whole web that you can run yourself, so a metasearch engine like this one will always pass your query on to other search engines.

You decide how far it goes:

  • Which search engines get your query: turn off the ones you do not want in Settings › How to calculate results › Engines.
  • Instant answers from outside services (weather, Wikipedia, Wolfram and others): the first switch under Settings › Advanced › Oracle & AI turns them all off.
  • Jev, the classifier that ranks, checks and tags results: "No Jev", in Settings › What to show on results, turns it and every other AI off, and greys out their options.
  • Search suggestions send what you type to a suggestion service; they have their own switch at the top of Settings.
  • Search in each of my languages, when you turn it on under Localization, sends your query to the translation service.

With No Jev on, outside instant answers off and suggestions off, your query goes to the search engines you chose and nowhere else. A few features send a result's address (never your query) from this server: site icons, the archive check, the TLS badge, and domain age (only the domain name, to its registry).

What each AI feature sends:

  • Plan searches sends searches of two words or more, never a bang, quoted search, address or secret, with nothing about you: to Jev up to four times (once to decide, and once per planned search to check it hasn't drifted) plus one model call; a plan is cached for a day to TypeSafe (the Jev classifier), then DeepInfra (DeepSeek V4.1 Flash) (their privacy policy). To turn off just this: the "Plan searches" box in How to calculate results, No Jev, or Use no outside services.
  • Plan while typing sends for this server's owner only: the text typed so far into the search box, when they pause on three words or more, up to three times a search, even if they then delete it or don't search: to Jev twice each time, and to the model at most twice. Only measured, with no words kept; results don't change to TypeSafe (the Jev classifier), then DeepInfra (DeepSeek V4.1 Flash) (their privacy policy). To turn off just this: the owner's "Plan while typing" box, Plan searches, Suggestions off, No Jev, or Use no outside services.
  • Translation sends the text being translated, and your query when you search in each of your languages to DeepInfra (DeepSeek V4.1 Flash) (their privacy policy). To turn off just this: No Jev in What to show on results (translation then goes to the translation services instead).
  • Which did you mean? sends queries of four words or fewer, with their first ten result titles to TypeSafe (the Jev classifier), then DeepInfra (DeepSeek V4.1 Flash) (their privacy policy). To turn off just this: the "Which did you mean?" box in What to show on results.
  • Go to the site a search names sends searches that start or end with a site's name ("youtube rust tutorial") to TypeSafe (the Jev classifier) (their privacy policy). To turn off just this: the "Go to the site a search names" box in What to show on results.
  • Dictionary that fits sends the query, when a dictionary card has several dictionaries to TypeSafe (the Jev classifier) (their privacy policy). To turn off just this: the "Open the dictionary that fits" box in What to show on results.
  • My algorithm's advice for engines sends your algorithm (once, when it names a search engine) and your query, for each engine it advises to DeepInfra (DeepSeek V4.1 Flash) (their privacy policy). To turn off just this: the "Follow my algorithm's advice for engines" box under My algorithm.
  • More like these sends your query and the text of the results you picked, when you press "Find more like these" to DeepInfra (DeepSeek V4.1 Flash) (their privacy policy). To turn off just this: the "More like these" box in What to show on results.
  • Research after results sends your query with the titles and snippets of your first ten results, then the searches the model proposes to DeepInfra (DeepSeek V4.1 Flash), then TypeSafe (the Jev classifier) (their privacy policy). To turn off just this: the "Research after results" box in What to show on results.
  • Follow-ups sends your query and the text of the first eight results' pages to TypeSafe (the Jev classifier), then DeepInfra (DeepSeek V4.1 Flash) (their privacy policy). To turn off just this: the "Follow-ups" box in What to show on results.
  • Page dates sends dates found on the first ten results' pages, with the words around them to TypeSafe (the Jev classifier) (their privacy policy). To turn off just this: the "Page dates" box in What to show on results.
  • Images, videos and maps in All sends the query to TypeSafe (the Jev classifier) (their privacy policy). To turn off just this: the "Images, videos and maps among All results" box in What to show on results.
  • Oracle routing sends the query, when no built-in pattern matched to TypeSafe (the Jev classifier) (their privacy policy). To turn off just this: the "AI oracle routing" and "Quick yes/no answers" boxes in Oracle & AI.
  • Result indicators sends the text of the first ten results' pages, and your own indicators to TypeSafe (the Jev classifier) (their privacy policy). To turn off just this: the "Result indicators" box in What to show on results.
  • Results that add nothing sends the text of the first ten results' pages that share a site or a title to TypeSafe (the Jev classifier) (their privacy policy). To turn off just this: the "Point out results that add nothing" box in What to show on results.
  • Original source of videos sends a video result's title and description to TypeSafe (the Jev classifier) (their privacy policy). To turn off just this: the "Original source for videos" box in What to show on results.
  • Reads on sends your query with the text and links of the first five results' pages to TypeSafe (the Jev classifier) (their privacy policy). To turn off just this: the "Reads on" box in What to show on results.
  • Rank by my algorithm sends your query; the titles, snippets and addresses of the results on the page and up to sixty below it; your "My algorithm" text; and the first 2,500 characters of each page to TypeSafe (the Jev classifier) (their privacy policy). To turn off just this: the "Rank by my algorithm" box in How to calculate results.
  • Describe images sends the thumbnails of the first 24 image results (not your query); descriptions are kept and shared to DeepInfra (Qwen3-VL 235B) (their privacy policy). To turn off just this: the "Describe images" box in What to show on results (deluxe; on by default).
  • Research when a search needs it sends your query with its first ten result titles to TypeSafe (the Jev classifier), then DeepInfra (DeepSeek V4.1 Flash) (their privacy policy). To turn off just this: the "Research when a search needs it" box in Oracle & AI.
  • AI research agent sends the query, the top results' pages it reads, and any attachment you gave it, when the research agent runs (in Lite it only decides which cards to show) to DeepInfra (deepseek-ai/DeepSeek-V4.1-Flash; keeps nothing) (their privacy policy). To turn off just this: "AI on search" and "AI on double-search" set to Off in Oracle & AI.

For agents and scripts

A script testing 4got should send the header X-4got-Test: 1 (or add test=1 to the address if it cannot set headers), so its searches stay out of Trending and the public log views; they are still logged for the owner, marked as a machine's. This is separate from opting out of logging.

A test search gets no Jev, no language model, no Wolfram and no browser sidecar. Name what you are testing in X-4got-Want: rerank,merge (or want=rerank,merge in the address); want=all gets everything. The names: routing, universal, indicators, source, merge, crawl, rerank, pagedates, platform_go, dict_pick, page_profile, cache_judge, planner, prefetch, clarify, followups, research_after, deep, research, translation, morelike, engine_advice, captions, each_language, wolfram, sidecar. The owner secret alone does not make a search a test. See the API.

Privacy

  • No IP logging
  • No user accounts or profiles
  • No tracking cookies — settings are the only cookies, and only non-default values are stored
  • Image proxy prevents tracking pixels from reaching you
  • TLS & ECH badge (if on): this server connects once to each result site, cached a month. Countries from a local copy of IP Geolocation by DB-IP (CC BY 4.0)
  • A "whois example.com" search asks the domain's registry
  • Malware warnings use a local copy of abuse.ch's URLhaus list; no address is sent out
  • This server monitors uptime of sites people report friction on: one request a minute while a site is down, then less and less often once it is back, never more than daily. Shift-click or hold a result's friction button to see a site's record
  • Privacy badges on every result: ToS;DR grades and tracker detection show you who's tracking what before you click
  • Result URLs from walled-garden sites (YouTube, Reddit, Twitter, etc.) are automatically rewritten to privacy-friendly alternatives (Invidious, Teddit, Nitter, etc.) — configurable in settings
  • Proof-of-work auth for returning visitors; no third-party CAPTCHAs
  • Tracking parameters are stripped from all result URLs

Features

  • 50+ oracle types — instant answers for weather, distances and travel times, stocks, crypto, currency conversion, dictionary, thesaurus, calculator, unit conversion, timezone, base conversion, color preview, package versions (npm/pip), DNS lookups, WHOIS, IP geolocation, SSL certificates, CIDR subnets, JWT decoding, cron expressions, DOI resolver, air quality, earthquakes, flight tracking, recipes, radio stations, sports scores, inflation/purchasing power, corporate ownership (SEC EDGAR + GLEIF), Wikipedia infobox, QR codes, UUID generation, password generation, roman numerals, chmod calculator, date arithmetic, hash/base64, unix timestamps, speed test, and more
  • AI research agent — a tool-using LLM that fetches top search result URLs, reads them, and synthesizes a comprehensive answer with citations, streaming findings in real time
  • Visual search with Lens — image upload, OCR text extraction, QR code detection, region selection for reverse image search
  • Federation — peer cache sharing, service proxying (translation, Wolfram, autocomplete), and instance discovery between friend instances
  • 15 search categories — web, images, videos, news, music, science, IT, files, maps, social, books, Q&A, packages, weather, lyrics
  • Bang redirects — !g, !w, !ddg and hundreds more for instant engine-specific search
  • RSS/Atom feeds — subscribe to any search query as an RSS feed, or subscribe to server news via Atom
  • Deluxe mode — some engines (Google, Qwant, Yandex) need a real browser; deluxe mode routes through a Playwright backend
  • SSH interface — search from the terminal via SSH

Oracles

Corporate ownership queries ("who owns X", "parent company of X") use SEC EDGAR (the US Securities and Exchange Commission filing database) and GLEIF (the Global Legal Entity Identifier Foundation registry) to trace parent companies, subsidiaries, and shareholder relationships.

VIN decoding ("vin …" or a valid 17-character VIN) uses NHTSA vPIC; only the VIN is sent.

Postal-code lookup ("zip 90210", "postal code 10115 germany", or a bare five-digit number) uses zippopotam.us, a free service over GeoNames data covering about sixty countries; only the code and the country go with the request.

AI research

The AI research agent activates when you search the same query twice (or click the Research button). It uses a tool-calling LLM to fetch and read the top search result URLs, extract key findings, and synthesize a comprehensive answer with citations. You see findings as they arrive.

Works with any OpenAI-compatible provider, including local Ollama.

Set 4got as your browser's default search engine.

This server

See the log page for available engines, engine health, and changelog. See the API documentation for programmatic access. Subscribe to server news via Atom. To follow a search, make it a standing search (Settings › Scheduled & standing searches) and use its private feed link: a feed reader then gets each new result it finds.

Source

Written in Go. Licensed under AGPL-3.0. Inspired by 4get and SearXNG.

Read the full documentation for detailed setup and configuration.

Send feedback