browser-search

browser-search

A skill for AI agents. OpenCode, Claude Code, Cursor, OpenClaw and beyond. Search the web with SearXNG, browse with Camofox, bypass protections with CloakBrowser. Anti-hallucination by design. All self-hosted, free, unlimited.

Why it exists

Today an AI agent browsing the web is like a thief in a balaclava sneaking around a police academy. The site protections block it, challenge it, reject it.

👮 browser-search flips the situation: your agent stops being the thief and becomes the chief of police. No more clumsy access attempts. It walks through every door because it has the right tools. SearXNG to search, Camofox to browse, CloakBrowser when the site plays hardball.

The skill enforces the exclusive use of deterministic scripts. This eliminates model hallucinations, even with the cheapest models. The 3 tools are described in natural language, but execution is rigid: the model can neither get the command wrong nor misinterpret the output. The result is guaranteed success on every query — the skill and its deterministic scripts guide the model to scour the web until it finds the answer.

1. SearXNG — metasearch engine for the search phase (multi-source, JSON)
2. Camofox — browser navigable via REST API for standard sites
3. CloakBrowser — stealth browser for anti-bot protected sites

The typical flow: the agent first searches with SearXNG, then browses the results with Camofox (or CloakBrowser if the site is protected).

Benefits

• 100% free, self-hosted, unlimited. No API keys to buy, no subscriptions, no rate limits. Everything runs on your machine, Docker and npm. Unlimited usage, zero cost.

• Lightweight, runs anywhere. Built and tested on a Raspberry Pi — if it runs there, it runs everywhere. Minimal resource consumption, no heavy infrastructure needed, runs 24/7 on low-power hardware.

• Search + browse in one kit. No manual integration needed. Searching and browsing are two distinct phases, both covered.

• Automatic navigation escalation. If Camofox gets blocked by Cloudflare/Akamai, the agent automatically switches to CloakBrowser.

• Smart performance. SearXNG for the search phase (milliseconds). Camofox and CloakBrowser are only used to browse the sites that actually need it.

• Automatic agent choice. The AI agent decides which tool to use: SearXNG for initial search, Camofox for browsing, CloakBrowser if the site is protected. Zero human intervention.

• Anti-hallucination by design. The skill's Deep Research mode enforces a "search first, answer second" workflow: the agent must verify every factual claim against live web sources, cross-reference multiple angles, and never guess. No more made-up answers.

• Fully customizable. The SKILL.md is plain text. You can edit the core rules, add your own, remove what you don't need. Adapt it to your workflow, your team, your standards.

• Native stealth. CloakBrowser automatically detects Cloudflare, Akamai, DataDome, Imperva, PerimeterX, and DDoS-Guard challenges, and waits for them to resolve before extracting content.

• Works with any agent. The SKILL.md is written for OpenCode, but the logic is identical for any AI agent. Same README, same package.json, everything works everywhere. Just ask your agent how to convert the skill for its environment.

Why it's different

Most web tools for AI agents stop where the wall starts: when a site throws a Cloudflare or Akamai challenge, they hand back an error and the agent gives up. browser-search is built to walk through instead.

• Search → browse → bypass in one skill. No manual integration, no per-query costs, no API keys to buy.
• Three-tier escalation. SearXNG for milliseconds-fast search, Camofox for the ~90% of standard sites, CloakBrowser when the site plays hardball — automatically, with zero human intervention.
• Proven anti-bot results. CloakBrowser applies 58 C++ source-level patches and scores 0.9 on reCAPTCHA v3 (human-level, server-verified), clearing Cloudflare, Turnstile, DataDome, Akamai, Imperva, PerimeterX and DDoS-Guard — where other tools just return blocked_by_challenge.
• Free, private, self-hosted. SearXNG, Camofox and CloakBrowser all run on your machine. Zero cost, unlimited, nothing metered.

Where other tools give up, browser-search keeps going.

🏆 State of the art

These three tools were chosen because they represent the current state of the art available today. A skill like this is designed to evolve: when better tools emerge, updating the SKILL.md is all it takes to swap them in. 🔄

⭐ Star the repo and follow to stay up to date on new tools, flow improvements, and orchestration updates over time. 🚀

Architecture

┌─────────────────────────────────────────────────────────┐
│                    browser-search                       │
│                                                         │
│  ┌──────────────┐                                       │
│  │    Search    │                                       │
│  │              │                                       │
│  │  SearXNG     │   search engines → URLs               │
│  │  (Docker)    │   JSON results, fast                  │
│  │  :8080       │                                       │
│  └──────────────┘                                       │
│         │                                               │
│         │ results ready → to browse                     │
│         ↓                                               │
│  ┌─────────────────────────────────────┐                │
│  │           Browsing                  │                │
│  │                                     │                │
│  │  ┌──────────────┐                   │                │
│  │  │   Camofox    │  browser + REST   │                │
│  │  │  (Docker)    │  JS, click, eval  │                │
│  │  │  :9377       │                   │                │
│  │  └──────┬───────┘                   │                │
│  │         │                           │                │
│  │         │ if blocked                │                │
│  │         ↓                           │                │
│  │  ┌──────────────┐                   │                │
│  │  │ CloakBrowser │  stealth Chromium │                │
│  │  │   (npm)      │  anti-bot, proxy  │                │
│  │  └──────────────┘                   │                │
│  └─────────────────────────────────────┘                │
└─────────────────────────────────────────────────────────┘

How it works

Phase 1 — Search with SearXNG

Docker container on localhost:8080. Metasearch engine that queries Google, Wikipedia, Bing, DuckDuckGo and many others simultaneously. JSON output with titles, snippets, and URLs.

Example:

node scripts/searxng/searxng.mjs search "largest llm benchmark 2026"

The agent now has a list of URLs to visit and autonomously decides whether to browse them with Camofox or CloakBrowser based on the site.

Phase 2 — Browse with Camofox

Docker container on localhost:9377. Exposes a full Firefox browser through a REST API. The agent can create tabs, navigate, click, scroll, execute arbitrary JavaScript, and structure data.

Includes: Mozilla's Readability.js for extracting clean articles, removing nav, sidebar, and ads (~70% token savings).

Main commands:

# Single-URL extraction (Readability.js, auto-fallback to snapshot)
node scripts/camofox/camofox.mjs readability "https://example.com"

# JavaScript evaluation
node scripts/camofox/camofox.mjs evaluate "https://example.com" "document.title"

# Accessibility snapshot
node scripts/camofox/camofox.mjs snapshot "https://example.com"

Phase 3 — Browse with CloakBrowser (when Camofox isn't enough)

npm package based on Playwright + cloakbrowser. Launches a Chromium browser with advanced fingerprinting to bypass Cloudflare, Akamai, DataDome and other anti-bot systems. Automatic challenge detection with wait and retry.

Available scripts:

• cloak-fetch.mjs — universal fetch with challenge detection
• cloak-script.mjs — custom Playwright script execution

Example:

node scripts/cloak/cloak-fetch.mjs "https://protected-site.com"
node scripts/cloak/cloak-fetch.mjs "https://protected-site.com" --proxy socks5://... --geoip

# Markdown output (requires: pip install markitdown)
node scripts/cloak/cloak-fetch.mjs "https://example.com" --format markdown

Version note: browser-search pins CloakBrowser ^0.5.5 and Playwright Core 1.62.1. CloakBrowser 0.5.x adds Linux arm64 support (Raspberry Pi) and Windows fixes. Known issue: on Windows with a Pro license, the browser may exit ~10s after launch (#479) — workaround: --license-through-proxy.

Why both Camofox and CloakBrowser?

Because speed and stealth are a tradeoff, and the right tool depends on the site.

Camofox — fast, structured, persistent. Camofox wraps Camoufox (a C++-level Firefox fork) in a REST API with an always-warm browser. After a ~1-3s cold start, every request is near-instant. Its accessibility snapshots are ~90% smaller than raw HTML, with stable element refs (e1, e2, ...) for reliable interaction. It handles the ~90% of sites that don't use advanced anti-bot protection: articles, docs, search engines, standard web pages.

CloakBrowser — stealth, anti-bot, on-demand. CloakBrowser launches a fresh Chromium instance per request (~1-3s startup each time). It uses advanced fingerprinting, proxy support, geoip, and automatic challenge detection to bypass Cloudflare, Akamai, DataDome, Imperva, PerimeterX, and DDoS-Guard. It is the last resort for the ~10% of sites that block Camofox.

Real-world numbers: