Here’s the thing: browser-search by Johell1NS is a skill suite that plugs into Claude Code, Codex, OpenCode, or any major coding agent and gives it real web search and page-reading abilities. Self-hosted. Zero API fees. No data leaving your infrastructure.

Three Components, One Pipeline

So here’s how the stack works:

• SearXNG — the meta-search engine that aggregates results from Google, Bing, DuckDuckGo, and others. Single endpoint your agent calls to search the web.
• Camofox — cloud-based browser renderer that reads full page content. Not just snippets like a search API — actual rendered text from the page your agent needs to read.
• CloakBrowser — open-source anti-detection layer that prevents cloud IDE sessions from getting blocked during automated browsing.

Still, the real win is how they form a search → browse → read pipeline your agent controls from start to finish, without ever touching a third-party AI search API.

Quick Install & First Test

And installing it was dead simple. I ran npm install -g @johell1ns/browser-search on my Ryzen 9 workstation — took about 25 seconds. Then configured the SearXNG endpoint (pointed at a public instance first for quick testing), dropped in the Camofox API key, then registered the skill in my Claude Code agent config. And just like that — from zero to “agent can search the web” in under 10 minutes.

Then I ran the real test. So I asked my agent: “Search for the latest AI agent frameworks released in June 2026 and compare their GitHub star growth.” And it hit SearXNG, browsed three result pages through Camofox, READ the actual READMEs, and came back with grounded data — project names, star counts, release dates. No “I think” or “based on my training data.” Real web content, real answer.

That’s the anti-hallucination play. But your agent isn’t guessing from its training cutoff — it’s reading the live web and reporting what it finds.

browser-search + Peerd: Complementary Tools

But one thing I want to highlight: browser-search is complementary to Peerd (which I covered earlier today). And Peerd lets your agent control a browser to interact with web apps — fill forms, click buttons, browse sites. browser-search gives your agent its own search+browse ability. Two different parts of the agent-browser spectrum, and they work together nicely.

Why Self-Host on a VPS

Now, for the full private setup, you’ll want to self-host SearXNG on a VPS. A basic DigitalOcean or Vultr droplet handles the Docker Compose deployment easily — SearXNG runs in a container with minimal resources. Public instances exist and work fine for testing, but running your own means your agent’s search queries stay on hardware you control. No third party sees what your agent is researching.

So that’s the data sovereignty argument, and it’s a strong one for production agent workflows. If you’re setting up a VPS for the first time, my Hermes VPS deployment guide covers the basics of getting a droplet ready for self-hosted tools.

Caveats & Limitations

But there are a few things to watch out for. But three components means three things to configure. It’s not a single-package-and-done situation — you need a SearXNG endpoint and a Camofox API key. Sure, the free tier of Camofox is generous but has limits. CloakBrowser is mainly useful if your agents run in cloud IDE environments like GitHub Codespaces; skip it if you’re running agents locally.

Also worth noting: the project is relatively young — 208 stars, last commit three days ago. The core works, but expect some rough edges in the config flow, especially around the CloakBrowser integration.

Bottom line. If you’re running coding agents and want them to stop hallucinating, browser-search is the most practical self-hosted solution I’ve found. It’s free, composable across agent frameworks, and the SearXNG integration gives you complete control over your agent’s search pipeline. The VPS angle for self-hosting is a natural fit if you want the full private setup.