Your AI agent’s web browser, minus the API keys and billing

What it does Hound is a single local MCP server that hands AI agents six web tools: fetching, crawling, keyless search, screenshots, cache management, and version checks. It runs entirely on your machine, queries ten public search backends in parallel, and automatically escalates blocked fetches to a stealth browser that patches its own fingerprints and simulates human mouse movements. The idea is to let an agent research the web without shipping data to a third-party scraper or signing up for another SaaS.

The interesting bit Instead of begging for API keys, Hound scrapes ten public search engines locally, deduplicates the results, and reranks them with a local ONNX cross-encoder model; it even derives follow-up queries from result snippets without calling an LLM. When a site throws up Cloudflare or DataDome, it does not give up—it swaps in a real Chrome TLS fingerprint, adds Bezier-curve mouse wiggles, and tries again.

Key highlights

• smart_fetch tries a fast HTTP request first, then auto-escalates to an anti-detect Patchright browser with Cloudflare Turnstile solving and JS-layer patches.
• smart_search queries ten keyless backends—DuckDuckGo, Brave, Mojeek, Yahoo, Yandex, Startpage, Google, Qwant, plus optional Wikipedia and Grokipedia—in parallel, then merges and neural-reranks results locally.
• Every response includes structured metadata—content_ok, page_type, relevance_score, engines_consensus, is_stale—so agents can branch on fields instead of parsing prose.
• Memory is tightly capped: a single renderer process, 512MB V8 heap limit, and CDP-triggered garbage collection after each fetch to keep RSS flat.
• Lean installs skip the browser entirely; on Termux or aarch64 it degrades gracefully to HTTP-only mode.

Caveats

• Sustained keyless search from a single IP will eventually hit rate limits or CAPTCHAs; the README admits this is a “gray-area posture” shared with SearXNG and does not claim search-engine ToS compliance.
• The neural reranker downloads an ~80MB ONNX model on first use; lean installs fall back to simpler cross-engine consensus ranking.
• Heavy or commercial use realistically needs a proxy set via environment variables to avoid blocks.

Verdict Agents that need autonomous, offline-capable web research without subscription fees should look here. Developers who already pay for SerpAPI or ScrapingBee and need guaranteed SLAs will not find them in a free, keyless scraper.