Gsearch lets AI agents run real Google searches through a real browser

What Gsearch does

As github.com reported in a Show HN post, Gsearch is an open-source server that gives AI agents access to Google Search. It speaks MCP, the Model Context Protocol, a standard way of plugging outside tools into an AI assistant. An agent calls a tool named google_search and gets back Google's results page as structured, labeled data (JSON) rather than a wall of messy web code.

The output is more than a list of links. It includes the AI Overview, the "People also ask" questions and related searches. The project says it exposes two tools in total.

How it gets past Google's defenses

The project's pitch is that simple approaches fail. According to the README, a plain web request hits a JavaScript wall, and an automated Chrome browser was sent to a CAPTCHA (a test meant to stop bots) 2 out of 2 times in the authors' test.

Gsearch instead runs Camoufox, a hardened version of Firefox, and lets it load Google's page as it is. It uses a visible window, since the authors say headless sessions (browsers with no window) were challenged more often on proxy IPs. The query is typed into the address bar.

Google hides result links behind redirects. Gsearch follows each one to find the real destination address without loading the page itself. The README says it does not solve CAPTCHAs and does not forge Google's tokens. When it is blocked, it returns a clear error with a wait time instead of an empty list.

Setup, proxies and limits

Installation is a few commands, and a one-time setup step downloads the browser, about 1 GB. Then you register the tool with an MCP client such as Claude Code, Claude Desktop or Cursor, and ask in plain language, for example for the top three links on a topic.

Without a proxy, which is a middleman server that changes the IP address Google sees, Gsearch uses your own connection. It is capped at 20 searches an hour and 150 a day to protect your IP. The README reports about 1.2 seconds per search on a warm session with no CAPTCHAs in its testing.

For more volume you can add a proxy in many formats. Rotating residential proxies give a new IP per search, while static ones each get their own budget. A setting called GSEARCH_DATA_LIMIT_MB sets a hard cap on proxy data. A clean proxy IP costs about 0.6 MB and 14 s per search, but many IPs are already flagged by Google and cost more time. The README says static proxies have not been benchmarked yet.

The legal and safety caveats

The README is unusually blunt. It says automated access to Google Search may breach Google's Terms of Service, whatever the project's open-source license says. It points to a lawsuit, Google v. SerpApi, filed 2025-12-19, in which Google argues that getting past its SearchGuard challenge by automated means counts as circumvention under US copyright law. The authors say they run the challenge in a real browser and forge nothing, but note that the complaint also targets automated browsers.

It also advises checking your proxy provider's acceptable-use policy, since many forbid use on sites that ban automated access.

A separate risk is prompt injection: any web page in the results can contain text meant to steer your agent. Gsearch marks every result as untrusted content, but that only helps if your agent treats it that way.

How much to trust it

Our newsletter flagged this as a very new project, with 1 point and 0 comments on Show HN when listed. Most numbers in the README are the authors' own measurements, and we have not verified them. Cautious teams should wait for more evidence. Hobbyists and builders testing research agents are the natural audience.