<div align="center">

<img src="https://raw.githubusercontent.com/dondai1234/master-fetch/master/docs/hound-logo.png" alt="Hound logo" width="128">

🐕 Hound

Give your AI agent the web. $0. Two commands. No keys.

Fetch · crawl · bypass bot walls · read PDFs (even scanned) · search the web One MCP server · one warm browser · zero accounts · runs on your machine

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

Install · The 6 tools · Search · Comparison · Honest limits

</div>

<br>

<div align="center"> <img src="https://raw.githubusercontent.com/dondai1234/master-fetch/master/docs/hound-hero.png" alt="Hound gives your AI agent the web" width="860"> </div>

---

🎬 Demo

Same prompt, three tools. Hound does the whole thing on its own, search + fetch + crawl, locally. The others get stuck on the parts they don't do.

<div align="center"><b>OpenCode + Hound</b></div>

<video src="https://github.com/user-attachments/assets/4e69e476-9c46-40cd-9bdf-f1b45882ac3e" controls muted width="860"></video>

<div align="center"><b>Pi + Hound</b></div>

<video src="https://github.com/user-attachments/assets/18b502c0-1e7b-4cb3-a730-074a9fbd0993" controls muted width="860"></video>

---

✨ New in 11.1.5

Next-gen stealth engine. Hound now passes the hardest anti-bot targets.

• 🧬 Stealth engine: system Chrome auto-detection (channel=chrome for real TLS fingerprint), 4 coherent fingerprint profiles, JS-layer patches (HeadlessChrome UA fix, navigator.webdriver=undefined, canvas noise via getImageData+toDataURL interception, permissions API). Human behavior simulation (Bezier mouse curves, natural scroll, dwell time). CF Turnstile solver with human-like mouse movement.
• 📊 Real-world results: passes bot.sannysoft.com (all checks), bypasses Cloudflare Turnstile on CanadianInsider (hardest in a 31-site benchmark), Medium, StackOverflow, NowSecure, Glassdoor (DataDome). See the stealth benchmark below.
• 🧠 Memory optimized: --renderer-process-limit=1, V8 heap cap 512MB, Memory.simulatePressureNotification via CDP after each fetch (triggers Chrome GC). CDP session leak fixed. RSS stays flat across sequential fetches.
• 🐍 No more scrapling. All scrapling functionality replaced with hound's own modules: fetcher.py (primp-based HTTP), browser.py (patchright-based browser), extractor.py (trafilatura + markdownify). Smaller install, fewer transitive deps, faster cold start.
• 📱 Works everywhere. Lean install pip install hound-mcp pulls no browser deps. On platforms without playwright (Termux, aarch64), hound runs in HTTP-only mode with graceful degradation.
• 🔒 Fixed Response.css(): added cssselect as explicit dependency (was missing from lxml's dependency tree). Removed silent [] return on errors; invalid selectors now raise.
• 🔧 805 tests.

---

Why you should pick Hound

Hound is one MCP server that gives any agent (Claude Code, Cursor, OpenCode, Hermes, Pi, anything that speaks MCP) full web research from a single local process.

• 🆓 $0 forever, MIT: no keys, no accounts, no per-request billing, no data routed to a third-party scraper. Search is keyless and local.
• 🧠 Mastered on connect: a one-time instructions block hands the agent the mental model, the #1 workflow, and the known limits. Effective on turn one.
• 📐 ~2.7K tokens, 6 tools: hand-crafted tool defs, no Pydantic schema bloat. More capability than tools shipping 5K+.
• 🎯 Every response is actionable: content_ok, next_action, summary, page_type, content_age_days/is_stale, source_type/is_official, relevance_score, fetch_relevance. Agents branch on structured fields, not error text. Hard-blocks (404/bot/auth) return clean errors, not fake content.
• 🛡️ Production-safe startup + shutdown: cold start under 1s so the MCP handshake never times out; exits 0 with clean stderr, no crash-like teardown noise.

Hound is for the agent itself. You install it once; the agent calls it whenever it needs the web.

---

🚀 Quick start

pip install hound-mcp[all]          # fetch + crawl + keyless search + PDF + OCR + neural rerank
playwright install chromium         # the anti-detect browser engine

Then point any MCP client at the hound command. No arguments, no keys, no env vars. See Install for the lean option and Tell your agent to install it for a copy-paste prompt.

hound -v          # version + update status
hound -u          # update to latest (brick-proof, self-healing)
hound --doctor    # health check + fix advice
hound --rollback  # undo the last update

If hound ever breaks (a failed update, a locked launcher), recover with python ~/.hound/repair.py, or run hound --doctor to diagnose.

---

🧰 The 6 tools

Tool	One-liner
smart_fetch	Fetch any URL. HTTP first, auto-escalates to the anti-detect browser if blocked. Bulk, PDFs (with OCR + quality score), css_selector, focus, actions, pagination.
smart_crawl	Best-first same-domain crawl. Each page as markdown with content_ok + page_type (article / list / js_shell). discover_only, crawl_urls, focus, sitemap mode, time + token caps.
smart_search	Local keyless web search. 10 backends in parallel, merges + ranks with neural rerank + cross-backend consensus. relevance_score + engines_consensus per result.
screenshot	Capture a page as an image. For multimodal agents (canvas, image-of-text, visual layout).
cache_clear	Clear the fetch cache. all=true wipes everything.
version	Installed version + update status.

---

🔎 Local keyless search

<div align="center"> <img src="https://raw.githubusercontent.com/dondai1234/master-fetch/master/docs/hound-scene.png" alt="Hound fetches the web and brings it back to your agent" width="820"> </div>

No API key, no account, no third-party service. smart_search runs 10 keyless backends in parallel on your machine, merges, dedups, and ranks. It returns URLs + ranking, not page content: the agent smart_fetches whichever results match what it needs (the ranking is a hint, not a directive).

• 🌐 10 independent backends: duckduckgo, brave, mojeek, yahoo, yandex, startpage, google, qwant, plus opt-in wikipedia + grokipedia. Six+ independent index families, not the same feed twice.
• 🧠 Neural rerank: a local ONNX cross-encoder (ms-marco-MiniLM-L-6-v2, Apache-2.0) running on the onnxruntime Hound already ships for OCR. Exa-style semantic ranking, $0, on your machine. Model downloads once (~80MB, cached, not bundled). Lean installs fall back to cross-engine consensus + engine-position order.
• 🎯 Cross-backend consensus: a URL returned by several independent indexes gets a consensus boost: a free authority signal from merging, no extra fetches. Every result carries relevance_score (0–1), fetch_relevance (high/med/low), and engines_consensus.
• 🔍 find_similar: pass url=; Hound fetches a page you like, derives a query, and reranks candidates against that source page. Exa's find-similar, local.
• 🛡️ Never dead: a diversity quorum waits for at least 3 backends to contribute before returning, so a single backend's bias or rate-limit can't dominate. A backend that CAPTCHAs or rate-limits is circuit-broken for 60s and carried by the others. engine_blocked in the response reports which ones cooled down.
• 📊 Filters: site / exclude_sites (domain include/exclude), location / language / region (geo), page (0–10), freshness (day | week | month | year). Default 6 results. A quality filter drops low-relevance results instead of padding to the max with garbage.
• 📈 related_queries: follow-up queries mined from result titles + snippets (no LLM). Search one to refine a broad query.

Search is 100% HTTP: it never touches the browser (the single Patchright browser is smart_fetch's alone).

<details> <summary><b>🔧 Search Engine Resilience Layer</b></summary>

Scraping public engines from your IP can be rate-limited or CAPTCHA'd. No keyless local tool is bulletproof against sustained blocking without a proxy: Hound is honest about that, then makes the no-proxy case as reliable as possible for a single user:

#	Mechanism	What it does
1	Persistent warm session per engine	One long-lived session reused across searches: cookies + TLS accumulate, so the engine sees a returning human, not a fresh bot. Also faster.
2	Per-engine pacing + jitter	Within one search all engines fire in parallel (free); only same-engine bursts across searches get a small jittered delay.
3	Circuit breaker + cooldown	A blocked engine auto-cools (60s) while the others keep serving.
4	202 / 429 / 503 / 403 + Retry-After	DDG's HTTP 202 soft rate-limit is detected; Retry-After honored.
5	Fingerprint rotation	A pool of real Chrome / Edge / Firefox / Safari TLS profiles, picked per request.
6	Diverse pool + consensus	10 backends across 6+ index families run in parallel: no single engine is a bottleneck, and agreement across independent indexes is a free authority signal.
7	HOUND_SEARCH_PROXY	Route all engine requests through your own rotating / residential proxy: the bulletproof path for heavy use.

Same gray-area posture as SearXNG / ddgs; no search-engine ToS compliance is claimed. </details>

---

🌐 Fetch & anti-bot

smart_fetch tries plain HTTP first (~1s). If the site blocks HTTP or serves a JS shell, it auto-escalates to a Patchright anti-detect browser with Cloudflare challenge solving. Two tiers, nothing to configure.

• 🛡️ Built-in Cloudflare bypass: a single stealthy Chrome warms at startup. It closes after 5 min of idleness to free RAM (HOUND_BROWSER_IDLE_TIMEOUT, set 0 to keep it alive forever) and relaunches in ~2s on the next fetch. Pages close after each fetch, idle memory stays near baseline. One browser total.
• 🧬 Stealth engine (v11.1+): system Chrome auto-detection (channel=chrome for real TLS fingerprint), coherent fingerprint profiles, JS-layer patches (HeadlessChrome UA fix, navigator.webdriver=undefined, canvas noise via getImageData+toDataURL interception, permissions API), human behavior simulation (Bezier mouse curves, natural scroll, dwell time), and a Cloudflare Turnstile solver with human-like mouse movement. See the stealth benchmark below.
• 🎯 Query-focused extraction: smart_fetch(url, focus="...") returns only the BM25-relevant blocks. Cuts context 80%+ on long pages, no re-fetch (runs post-cache). Re-pass the same focus when paginating.
• 🖱️ Page interaction: actions=[{click:'button.load-more'},{fill:{selector:'#q',text:'x'}},{press:'Enter'},{wait:500},{scroll:3},{wait_selector:'.item'}] for load-more, search forms, pagination, infinite scroll. Forces stealthy + bypasses cache.
• 🏷️ Metadata on every response: title, description, site name, type, image, canonical URL, language, published time, author (OpenGraph + JSON-LD + canonical).
• 🔗 Outgoing links: include_links=true populates response.links classified as citations (main-content references, the ones worth following) / navigation / external + a primary_source hint. Follow a page's source chain in one step.
• 🐕 Reddit, optimized: Reddit URLs auto-rewrite to old.reddit.com (7× smaller) and skip to the stealthy browser. Subreddit listings parse into structured posts with promoted ads filtered out.
• 💾 Smart caching: SQLite (WAL mode), keyed by URL + extraction type + css_selector + pages. Bad content is never cached; a size cap evicts the oldest so a long-lived agent's cache can't grow unbounded. cache_ttl=0 forces fresh.
• 📐 Pagination: content over 40KB is chunked; the response gives next_offset so the agent pages through with one more call (served instantly from cache).

🧬 Stealth benchmark

Real-world results from v11.1.0, tested against hard anti-bot targets.

Detection test sites (all checks pass):

Site	What it tests	Result
bot.sannysoft.com	HeadlessChrome UA, webdriver, plugins, WebGL, permissions, Selenium	ALL PASS
CreepJS	Canvas hash, audio fingerprint, lie detection	200 OK
BrowserScan	CDP detection, fingerprint analysis	200 OK
Pixelscan	Headless detection, fingerprint consistency	200 OK

Anti-bot protected sites (content extracted):

Site	Protection	Status	Content
CanadianInsider	Cloudflare Turnstile (hardest in 31-site benchmark)	200	78 KB, title: "Canadian Insider"
Medium	Cloudflare interstitial	200	93 KB, title: "Medium"
StackOverflow	Cloudflare	200	1.1 MB, full question page
NowSecure	Cloudflare challenge	200	180 KB, title: "nowsecure.nl"
Glassdoor	DataDome	200	849 KB
Reddit	Cloudflare lite	200	1 MB
Hacker News	None (baseline)	200	35 KB, title: "Hacker News"
GitHub	None (baseline)	200	523 KB

Note: Google Search returns 429 (rate limited) as it uses its own bot detection independent of Cloudflare. This is expected.

Stealth signals verified:

Signal	Value	Detection status
navigator.webdriver	undefined (patched from false)	Not detected
navigator.userAgent	Chrome/150 (HeadlessChrome removed)	Not detected
navigator.platform	Win32 (real system Chrome)	Not detected
navigator.plugins.length	5 (real system Chrome)	Not detected
window.chrome	object (present)	Not detected
WebGL vendor/renderer	Real GPU (Intel UHD, Direct3D11)	Not detected
Canvas fingerprint	Per-session noise (different each session)	Not detected
TLS fingerprint	Real Chrome 150 JA4 (system Chrome)	Not detected

Memory efficiency (5 sequential fetches):

RSS decreased by 3.5 MB over 5 fetches. No RAM creep. The Memory.simulatePressureNotification CDP command triggers Chrome's internal GC + cache drop after each fetch (~5ms, non-disruptive).

---

🕷️ Crawl

smart_crawl walks same-domain links in best-first order: discovered URLs are scored by focus relevance + content-likelihood (docs/guide/api boosted, login/submit/cart penalized) + shallow depth, so content pages are crawled before junk when the budget is tight.

• 🎯 Content-adaptive extraction: article/docs → trafilatura main content; list/index pages (HN, aggregators, directories) → a structured * [title](url) link list; JS shells → detected and reported honestly.
• 🗺️ Sitemap mode: options sitemap=true maps the whole site from sitemap.xml in ONE fetch (full URL list + lastmod, no BFS). sitemap='auto' uses it if the site has one, else falls back to BFS. Collapses a hundreds-of-pages discovery crawl into one call.
• 📍 discover_only=true: URL map only (BFS-based). For big sites prefer sitemap=true instead.
• 🎯 focus='query': prioritizes relevant pages within the budget AND focus-filters each page's content.
• 📋 crawl_urls=[...]: second-phase selective crawl of a chosen subset (no re-discovery).
• 🛡️ Dedup + scoping: URLs normalized so /docs and /docs/ are never crawled twice. Same-domain only by default; path_include / path_exclude to scope.
• ⏱️ Caps: max_pages (default 10), max_depth (default 2), max_total_chars (token budget), deadline_ms (overall time, default 120000). Each page carries content_ok + status + fetched_at; next_action tells you if the crawl stopped early.

---

📄 PDF + scanned-PDF OCR

smart_fetch detects a PDF (by content-type or %PDF magic bytes) and extracts it to structured markdown with pdfplumber (MIT): multi-column reading order, real tables as markdown tables, font-size headings, de-hyphenated paragraphs, a metadata header, and --- Page N --- markers.

• 📑 table_of_contents: the PDF outline as [{level, title, page, end_page}]. PDFs without bookmarks get a heading-based fallback map. Pass pages='23-31' to grab one section by range and save tokens.
• 🔍 CID-corruption auto-OCR (the flagship trick): academic papers embed font subsets without a Unicode map, so extractors emit (cid:71)(cid:302)... garbage for figures/diagrams/math. But the glyphs render correctly. Hound detects CID-garbage pages, renders them via pypdfium2, and OCRs them with rapidocr, recovering the real text automatically.
• 🖼️ Scanned / image-only PDFs (and image-only web pages) are auto-OCR'd too. Pure-pip, no system binary, with [all].
• 📊 quality_score (0.0–1.0) + honest content_ok: trust PDF content more the closer the score is to 1.0.
• 📎 password for encrypted PDFs; include_media=true for per-page image metadata; a .pdf URL that returns a login/paywall is reported as auth_required.

---

📸 Screenshot

screenshot captures a page as an image. For multimodal agents only: use when content is rendered as images / canvas / image-of-text or you need visual layout. Text-only agents should use smart_fetch instead. A stealthy browser session is auto-managed.

---

📊 Comparison: free tools

Most free web tools for agents do one thing and miss the rest. Hound is the only one that bolts all of it onto a single local MCP server for $0, no keys.

	Hound	Crawl4AI	Parallel Search	Jina Reader	Firecrawl (OSS/free)
Price	$0 forever	$0 (self-host)	free, rate-limited	free, rate-limited	$0 self-host / 1K free
Runs locally	yes	yes	no (their servers)	no (their API)	self-host: yes (Redis + Docker)
Web search	yes (keyless local, 10 backends)	no	yes (remote)	yes	no
Deep crawl	yes (best-first, sitemap, budget)	yes	no	no	yes (cloud)
Anti-bot / Cloudflare	built-in (Patchright)	limited	yes (their infra)	none	not by default
PDF → structured markdown	yes (tables, ToC, subset)	partial	no	yes (native)	yes (cloud + OCR)
Scanned-PDF / image OCR	yes (rapidocr, pure-pip)	no	no	no	yes (cloud paid)
Page interaction	yes (actions)	hooks (code)	no	no	yes (cloud)
Query-focused extraction	yes (focus, BM25)	yes (BM25 filter)	no	no	no
Agent signals	yes (content_ok/next_action/summary/relevance_score)	no	no	no	no
Connect-time instructions	yes	no	no	no	no
MCP server	yes (official)	community	yes (official)	yes (official)	build it
Token cost (tools/list)	~2.7K (6 tools)	varies	n/a	n/a	varies (12 tools)

The short version: Crawl4AI crawls well but has no search and trips on Cloudflare. Parallel Search is remote search-only, no crawl, and runs on their servers. Jina fetches but rate-limits and routes through Jina. Firecrawl keeps the good stuff behind the paid cloud. Hound is the only free tool that combines keyless local search, built-in Cloudflare bypass, best-first crawl, scanned-PDF OCR, page interaction, and query-focused extraction in one local MIT server: $0, no accounts, no keys.

<details> <summary><b>When a paid service makes sense</b></summary>

Paid scrapers (Bright Data, ZenRows, Firecrawl paid, Spider.cloud) can beat free tools on the hardest anti-bot (DataDome, Akamai, Cloudflare Turnstile) and on massive scale, because they run large residential-proxy networks. Paid search APIs (Exa, Tavily) offer hosted neural search. They cost $16 to $500+/month, require accounts + API keys, and send your queries + content through their servers. Use Hound for $0 local web research with no accounts and no keys; reach for a paid service only for enterprise scale, sites Hound explicitly can't crack, or hosted neural search at scale. </details>

---

📦 Install

pip install hound-mcp[all]          # recommended: fetch + crawl + keyless search + PDF + OCR + neural rerank
playwright install chromium

<details> <summary><b>Lean install (HTTP-only mode, no browser/OCR)</b></summary>

pip install hound-mcp               # fetch + crawl + keyless search (HTTP-only, no stealthy browser)

The lean install works on all platforms (including Termux/Android). It gives you multi-engine keyless search, HTTP fetch with auto-escalation (HTTP tier only, no stealthy browser), crawl, and caching. Stealthy browser escalation and screenshot require browser deps from the [all] extra. </details>

<details> <summary><b>Optional environment variables</b></summary>

Variable	Purpose
HOUND_SEARCH_PROXY	Route all search-engine requests through your own proxy (http://host:port, socks5://..., or user:pass@host:port). For sustained heavy search use with a rotating / residential proxy. Not required for normal single-user use.
HOUND_SEARCH_MIN_INTERVAL	Override the per-engine pacing floor (seconds, float). 0 = use the built-in defaults (DDG 1.2s, Bing 1.5s, Wikipedia 0.3s). Power-user tuning.
HOUND_BROWSER_IDLE_TIMEOUT	Seconds of browser idleness before the warm Chrome is closed entirely to free RAM (default 300, i.e. 5 min). The next fetch relaunches it in ~2s. Set to 0 to keep Chrome alive forever (old behavior).

No API keys or accounts are needed for anything: search is keyless and local. </details>

<details> <summary><b>Updating, rolling back, and repairing</b></summary>

hound -u          # update to latest (brick-proof: --no-deps, detached helper, self-heal)
hound --doctor    # health check: launcher, imports, metadata, deps, PyPI, repair script
hound --rollback  # reinstall the version from before the last update

hound -u is designed to never brick the install. It updates with --no-deps (no heavy extras that fail mid-install); on Windows it runs pip in a detached helper after the launcher exits (Windows can't overwrite a running .exe), freeing the launcher via the rename trick. If a pip pass leaves the version unchanged, it self-heals with a --force-reinstall --no-deps pass.

If hound is ever broken (a failed manual pip while a server held the launcher, a half-finished update), the safety net is a standalone script written outside site-packages on every update:

python ~/.hound/repair.py   # stops hound, force-reinstalls hound-mcp from PyPI, verifies

It survives because it is not part of the hound-mcp package, so a failed pip uninstall never removes it. hound --doctor diagnoses the install and tells you the right fix. </details>

---

🤖 Tell your agent to install it

Paste this into your agent:

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 (OpenCode, Hermes, Pi, etc). Then find: (a) where the MCP config file lives, and (b) what format it expects for adding a local MCP server. Read the harness docs if needed. Do not guess.

2. Run: pip install hound-mcp[all]
   Then run: playwright install chromium (But only if it isnt installed already, verify first about its existence)
   If either fails, stop and tell the user.

3. Find the MCP config file from step 1 and back it up before editing. Add a new MCP server named "hound" with command "hound", no arguments, in the format your harness requires. No API keys or environment variables are needed (search is keyless and local).

4. Save the file. Tell the user to restart the agent. After restart, smart_fetch, smart_crawl, smart_search, screenshot, cache_clear and version should be available.

<details> <summary><b>For Pi agent users</b></summary>

Install the Hound MCP server, then the Pi extension:

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

No API keys, no config file, no MCP adapter needed. The extension spawns hound as a singleton subprocess and registers all 6 tools (web_fetch, web_search, web_crawl, web_screenshot, cache_clear, hound_version) as native Pi tools. Prewarmed at session start. Run /reload to activate.

Updating:

hound -u                              # update the MCP server
pi update npm:@houndmcp/hound-mcp-pi  # update the extension

The extension checks version sync at session start and warns if the extension and hound diverge by a major version.

</details>

<details> <summary><b>For Open WebUI (HTTP) users</b></summary>

Open WebUI v0.6.31+ speaks the streamable HTTP transport natively. Run Hound in HTTP mode and point Open WebUI at it, no mcpo proxy needed:

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

Then in Open WebUI add an MCP server with URL http://127.0.0.1:8765/mcp. Stdio clients (Claude Code, Cursor, OpenCode, Pi, etc.) just use hound with no flag.

</details>

---

⚠️ Honest limits

No free tool can do everything. Hound is upfront about what it can't:

Limit	What happens instead
DataDome / Akamai / Cloudflare Turnstile (interactive)	Not bypassed. next_action tells the agent to switch sources instead of retrying.
Search rate-limits / CAPTCHAs	Solved by diversity: 10 keyless backends run in parallel; a backend that rate-limits/CAPTCHAs is carried by the others, and a diversity quorum waits for 3 to contribute so no single backend dominates. Search is never dead. engine_blocked reports cooled-down backends; HOUND_SEARCH_PROXY is a power-user rotating-proxy escape hatch for per-IP throttling (the one thing no scraper can escape from one IP).
Neural / find_similar search	Need hound-mcp[all] (the ONNX reranker runs on the same onnxruntime as OCR; model downloads once). Lean installs get cross-backend consensus + engine-position ranking.
Sites requiring login	Out of scope (Hound does page interaction, not authenticated sessions).
Deep shadow-DOM / hard SPAs	actions (scroll, click, wait_selector) reach most of it; deep shadow-DOM piercing not yet wired.
YouTube	Minimal text.

When a fetch or search fails, the response says exactly why and what to try next, so the agent doesn't waste calls guessing.

---

🪙 Token cost

<div align="center"> </div>

Most MCP servers cost 3–5K tokens just to exist. Hound's 6 tools cost ~2.7K tokens at tools/list (measured with cl100k_base); the connect-time instructions (~0.8K, the orientation doc) are injected ONCE at handshake, not repeated every turn. Your context window is expensive; Hound respects it.

---

<div align="center">

If Hound saves you time, ⭐ the repo: it helps others find it.

MIT · Changelog · Issues · PyPI

</div>