What is Hound MCP server?

Hound is an open-source MCP (Model Context Protocol) server that gives AI agents web research capabilities. It provides six tools: fetch, crawl, search, screenshot, cache clearing, and version checking. It runs locally on your machine, costs $0, and requires no API keys. It works with any MCP-speaking agent including Claude Code, Cursor, OpenCode, Hermes, and Pi.

Can Hound bypass Cloudflare and anti-bot protection?

Yes. Hound uses system Chrome with auto-detection for real TLS fingerprints, four coherent fingerprint profiles, JS-layer patches for webdriver and HeadlessChrome, and human behavior simulation with Bezier mouse curves. It passes detection test sites like bot.sannysoft.com, creepjs, browserscan, and pixelscan. It bypasses Cloudflare Turnstile, CF Challenge, CF Lite, DataDome, and CF Interstice on sites like Medium, StackOverflow, Reddit, and Glassdoor. DataDome, Akamai, and interactive Cloudflare Turnstile are not bypassed.

How does Hound's keyless search work?

Hound searches 10 backends in parallel with no API key: DuckDuckGo, Brave, Mojeek, Yahoo, Yandex, Startpage, Google, Qwant, Wikipedia, and Grokipedia. Six independent index families. A local ONNX cross-encoder reranks results semantically. A diversity quorum waits for three backends before returning. Rate-limited backends are circuit-broken for 60 seconds and carried by the others.

How do I install Hound MCP server?

Run two commands: pip install hound-mcp[all] then playwright install chromium. For a lean HTTP-only install without browser dependencies, use pip install hound-mcp. Then add a new MCP server named hound with command hound and no arguments. No API keys needed. Works on Termux and aarch64 with the lean install.

Yes, completely free. Hound is open-source under the MIT license. It runs locally on your machine with no API keys, no accounts, no credits, and no rate limits beyond what the search backends themselves impose. 131+ stars on GitHub.

What can Hound not do?

DataDome, Akamai, and interactive Cloudflare Turnstile are not bypassed. The next_action field tells the agent to switch sources. Login-required sites are out of scope. Deep shadow-DOM piercing is not yet wired, though actions reach most of it. YouTube provides minimal text. Search rate-limits are handled by 10-backend diversity, not bypassed individually.

00 SURVEY

Hound MCP server — Give your AI agent the web.

One MCP server. $0. No keys, no accounts. Fetch, crawl, search, bypass Cloudflare, read PDFs. Runs on your machine.

pip install hound-mcp[all] && playwright install chromium

Get started View on GitHub

06 TOOLS / CAPABILITIES

Six tools. every web task.

smart_fetch

Fetch any URL. HTTP first, auto-escalates to the anti-detect browser if blocked. Bulk fetch, PDFs with OCR, page interactions, pagination.

auto-escalate

smart_crawl

Best-first same-domain crawl. Each page as markdown with content_ok and page_type. Sitemap mode, focus filtering, token budgets.

best-first

smart_search

Ten keyless search backends in parallel. Neural rerank with cross-backend consensus. No API key, no account.

10 backends

screenshot

Capture a page as an image. For multimodal agents that need visual layout.

multimodal

cache_clear

Clear the fetch cache. All or expired only. Set cache_ttl=0 to force fresh.

all / expired

version

Installed version and update status. Brick-proof self-update with rollback.

self-healing

08 SITES / STEALTH BENCHMARK

Passes the hardest anti-bot targets.

System Chrome auto-detection for real TLS fingerprints. Four coherent fingerprint profiles. JS-layer patches for webdriver, HeadlessChrome, canvas noise, permissions. Human behavior simulation with Bezier mouse curves. Cloudflare Turnstile solver with human-like mouse movement.

STEALTH TELEMETRY / LIVE READOUT

DETECTION TEST SITES

bot.sannysoft.com[PASS]
creepjs[PASS]
browserscan[PASS]
pixelscan[PASS]

ANTI-BOT PROTECTED SITES

SITE	PROTECTION	STATUS	SIZE
canadianinsider	CF TURNSTILE	200	78 KB
medium	CF INTERSTICE	200	93 KB
stackoverflow	CLOUDFLARE	200	1.1 MB
nowsecure	CF CHALLENGE	200	180 KB
glassdoor	DATADOME	200	849 KB
reddit	CF LITE	200	1 MB
hacker news	NONE	200	35 KB
github	NONE	200	523 KB

STEALTH SIGNALS

navigator.webdriver	undefined	not detected
navigator.userAgent	Chrome/150	not detected
navigator.platform	Win32	not detected
navigator.plugins	5	not detected
window.chrome	object	not detected
WebGL renderer	real GPU	not detected
canvas fingerprint	per-session	not detected
TLS fingerprint	Chrome 150	not detected

Memory: RSS decreased 3.5 MB over 5 sequential fetches. No RAM creep.

10 ENGINES / KEYLESS SEARCH

Ten engines. zero keys.

DuckDuckGo, Brave, Mojeek, Yahoo, Yandex, Startpage, Google, Qwant, plus Wikipedia and Grokipedia. Six independent index families, not the same feed twice. A local ONNX cross-encoder reranks results semantically. A diversity quorum waits for three backends before returning. A backend that rate-limits is circuit-broken for 60 seconds and carried by the others.

10 backends

6+ index families

$0 cost

01 CHOICE / WHY HOUND

The only free tool that does all of it.

Crawl4AI

Crawls well.

No search. Trips on Cloudflare.

Jina Reader

Fetches pages.

Rate-limits. Routes through their API.

Firecrawl

Managed scraping.

Good stuff behind paid cloud. Credits don't roll over.

Parallel Search

Remote search.

No crawl. Runs on their servers.

Hound

Fetch, crawl, search, bypass, PDF, OCR. All local. $0.

6 tools

~2.7K tokens

tools/list

Typical MCP servers dump 3-5K tokens just to exist.

02 COMMANDS / QUICK START

Two commands. Then point your agent at hound.

$shell

pip install hound-mcp[all]

$shell

playwright install chromium

01 Lean install (HTTP-only)

pip install hound-mcp

No browser deps. Works on Termux, aarch64. HTTP fetch + crawl + keyless search. No stealthy browser or screenshot.

02 Updating and rollback

hound -u — update
hound --doctor — health check
hound --rollback — undo last update
python ~/.hound/repair.py — emergency repair

03 Pi extension

pip install hound-mcp[all] then pi install npm:@houndmcp/hound-mcp-pi

All 6 tools as native Pi tools. Prewarmed at session start.

04 Open WebUI (HTTP)

hound --http --host 127.0.0.1 --port 8765

Point Open WebUI at http://127.0.0.1:8765/mcp

AGENT INSTALL PROMPT

Install the Hound MCP server on this machine. Follow every step. Do not skip any.
1. Figure out which agent harness you are running on. Find where the MCP config file lives and what format it expects.
2. Run: pip install hound-mcp[all]  Then: playwright install chromium (verify first if not already installed)
3. Add a new MCP server named "hound" with command "hound", no arguments. No API keys needed.
4. Save the file. Tell the user to restart the agent.

06 LIMITS / HONEST

What Hound can't do.

[BLOCKED]DataDome, Akamai, interactive Cloudflare Turnstile — not bypassed. next_action tells the agent to switch sources.
[RATE]Not an issue if you're a single user, even with decently heavy use. The 10 backends carry the load. HOUND_SEARCH_PROXY if you outgrow them.
[AUTH]Login-required sites — out of scope.
[DOM]Deep shadow-DOM — actions reach most of it. Deep shadow-DOM piercing not yet wired.
[VIDEO]YouTube — minimal text.