Hound MCP: Free Web Research for Your AI Agent

If you've ever asked Claude to research something and gotten a politely worded refusal — "I don't have access to the web" — you already understand the gap Hound fills. Your agent is capable. It's just missing the connection.

Hound is a model context protocol (MCP) server that hands your agent a working web browser and a multi-engine search function, installs locally in two commands, and costs nothing. No account creation, no API key to paste in, no monthly limit counting down in the background. It's been sitting quietly on GitHub since late 2025, picked up steady engagement in the MCP communities in mid-2026, and as of this writing the install story is about as low-friction as anything in the MCP ecosystem.

At a glance

What Hound actually is

A model context protocol (MCP) server is a small program that runs on your machine and exposes a set of tools to whatever AI agent connects to it. The agent doesn't know how to browse the web by itself — but when it's connected to an MCP server that does, it can call those tools as naturally as it uses any other capability. Claude Desktop, Cowork, Cursor, and most of the major agent interfaces now accept MCP servers through a short config entry.

Hound is that server for web access. Install it, point your agent's config at it, and your agent can fetch any URL, run a search across ten different search engines simultaneously, crawl an entire domain, and capture screenshots — all without you signing up for anything or entering a credit card number.

The zero-cost positioning is not a trial period. Hound is free because it doesn't route your requests through any paid service. Searches go through public, keyless backends — DuckDuckGo, Brave, Mojeek, Qwant, Yahoo, and others. Fetches use direct HTTP requests, with a full browser engine for pages that require JavaScript or actively block simple crawlers. Nothing phones home to a vendor.

What ships in the repo

Hound exposes six tools. The ones you'll actually use day-to-day are the first three.

smart_fetch is the core fetch tool. It tries a fast HTTP request first; if the page blocks it or requires JavaScript rendering, it escalates automatically to a stealth browser that mimics a real user. It handles PDFs — including scanned ones via OCR (optical character recognition) — and can simulate page interactions like clicking, scrolling, and filling form fields for sites that require them. When a page is blocked and no live version is available, it falls back to the Internet Archive and tells the agent exactly what it got and where from.

smart_search runs your query across ten search backends in parallel, re-ranks the combined results on your own machine (with the [all] install), and returns a single consensus list. The agent gets better signal than any single engine would return, and the whole thing runs on your hardware without any API call to an external service.

smart_crawl systematically reads every page within a single website, up to a limit you set — useful when your agent needs to read an entire documentation site, map a company's public pages, or pull all the articles from a news section. You set how many pages to crawl; it handles the rest.

screenshot captures a rendered image of any URL, which matters for agents that can read and reason about what a page actually looks like — not just its text.

cache_clear and version are housekeeping: clear the local page cache, or check whether your install is current.

Every response from smart_fetch and smart_search comes with structured metadata — whether the content was usable, what the page type was, and a suggested next step if something went wrong. When a fetch fails, the agent gets a reason and a recovery hint, rather than silence.

Why this matters for non-coders running agents

The population this is most useful for: anyone who has set up Claude Desktop or a similar agent and runs research-heavy workflows — competitive analysis, client due diligence, market sizing, literature reviews, news monitoring, sourcing vendors or candidates. These are not tasks that need a developer to configure; they need the agent to be able to browse.

Without Hound, that means either relying on your agent's training-data cutoff, or paying for a hosted search service. Tavily's free tier runs out. Firecrawl's free tier is rate-limited and requires a key. Brave Search API requires registration. Hound requires none of that. For someone who wants to install once and never think about per-call billing again, the calculus is simple.

What Hound does not cover: authenticated sites (CRMs, internal wikis, anything requiring login) and pages behind CAPTCHAs like Cloudflare Turnstile or DataDome. The use case is open-web research, and within that scope it's comprehensive.

Where it fits

The web-research plugin space has several options worth knowing:

mzxrai/mcp-webresearch — the early community favorite for giving Claude live web access. Three tools: Google search, page visit, screenshot. Archived in February 2025, no longer maintained. Still works, but you're on your own for updates.

Firecrawl MCP — the most capable web-data server in the ecosystem, covering search, scraping, crawling, browser automation, and an autonomous research agent in one install. Firecrawl's free tier allows some keyless usage, but serious use requires an API key and the costs scale with volume. If you need production-grade crawling at scale, Firecrawl is the call. If you want $0 forever for personal-research workloads, Hound wins.

Tavily — purpose-built search API for AI agents with good structured outputs. Requires an API key; free tier is limited. The quality of results is high, but you're metered.

Crawl4AI MCP — seventeen tools, full browser rendering for JavaScript-heavy pages, AI-powered summarization, and solid deep-crawl capabilities. More complex to configure than Hound; also free and self-hosted. The right choice if your primary need is bulk crawling with summarization rather than ad-hoc research.

Hound's position: the keyless, lowest-friction option for everyday agent research. Not the most powerful, not the most scalable — but the one that genuinely costs nothing to run forever.

Verdict

Install. For any non-coder running an agent who wants web access without a monthly tab, Hound is the obvious move. Two install commands, four lines of JSON, and six tools covering the full range of open-web research a knowledge worker needs. The maintenance cadence is healthy — v10.4.0 shipped the same day as this article.

The caveats are narrow: login-required pages are out of scope, interactive CAPTCHAs will sometimes block fetches, YouTube returns minimal text. If your workflow depends on any of those, Firecrawl or Tavily are better fits. For everything else, Hound earns a permanent slot in your agent config.

For your agent to install this

You can skip this section — it's the install details your AI agent needs. If you're using Cowork, Claude Code, Claude Desktop, ChatGPT Work, Codex, or Hermes, say "install hound MCP for me" and it will read the tables below.

System requirements

API keys required

None. No accounts, no keys, no environment variables needed for standard use.

Optional environment variables for advanced use:

• HOUND_SEARCH_PROXY — route search through a custom proxy
• HOUND_SEARCH_MIN_INTERVAL — override per-engine pacing in seconds
• HOUND_BROWSER_IDLE_TIMEOUT — browser close timeout in seconds (default 300)

Install commands

pip install hound-mcp[all]
playwright install chromium

The [all] extra includes neural reranking and PDF/OCR support. Lean install (omit [all]) works fine but skips those features.

Claude Desktop config (claude_desktop_config.json)

{
  "mcpServers": {
    "hound": {
      "command": "hound"
    }
  }
}

That's the complete config. No arguments, no environment block needed.

Health check

hound --doctor

Run this after install to verify the setup. hound -v reports the installed version and whether an update is available.