Novada MCP

One MCP server. All web data. Search, extract, crawl, scrape, proxy, and AI research — 22 curated tools behind a single npx command. Run it locally or call the hosted endpoint (15 tools).

Always use npx -y novada-mcp@latest. A bare npx novada-mcp or a global npm i -g novada-mcp can silently run an old cached version. See Staying on the latest version.

# Local (Claude Code)
claude mcp add novada -e NOVADA_API_KEY=your_key -- npx -y novada-mcp@latest

The Problem

AI agents need web data but the tools are fragmented:

• Tavily does search but can't scrape or proxy
• Firecrawl does scrape but can't search or proxy
• BrightData does everything but ships 69 tools that bloat your context window
• Building it yourself means maintaining proxies, anti-bot bypass, retry logic, and 10 different APIs

The Fix

npx -y novada-mcp@latest

One server. One API key. Tools that cover every web data need an AI agent has:

Tools at a Glance

22 curated tools, 6 categories (the count novada_discover shows). The hosted endpoint exposes a 15-tool subset. Older tool names (e.g. novada_health, novada_unblock, novada_verify, the six typed novada_proxy_*, wallet/traffic lookups) still work as back-compat aliases that route into these. Load all, or scope with NOVADA_GROUPS / NOVADA_TOOLS.

Don't want all 22 in your context window? NOVADA_GROUPS="search,extract" loads only those groups. Full list any time via novada_discover.

What Makes This Different

novada_research is unique. No other MCP server turns one question into ranked, cited source material in a single call. It searches across Google, Bing, and DuckDuckGo in parallel, deduplicates, extracts full content from the top sources, and returns the most relevant passages under numbered source sections for your agent to reason over (extractive, not a generated prose report — you compose the answer). One tool call replaces an entire search→extract workflow. Depth options: quick (3 queries), deep (6), comprehensive (8-9).

Auto-escalation handles anti-bot automatically. Static fetch → JS render → Browser CDP. Known hard targets (Amazon, LinkedIn, G2, Zillow, Glassdoor, Walmart, Instagram, TikTok, Shein) skip straight to the right method based on a 30+ domain registry. You never think about Cloudflare, DataDome, Kasada, or PerimeterX — the tool handles it.

Agent-first design (8.5/10 benchmark score). Every response includes agent_instruction with structured next-step guidance, source field (live/cache/wayback), structured errors with failure_class, cross-tool hints suggesting better alternatives, and a ## Agent Action block with machine-parseable status codes.

Tool Reference

Search & Research

Extract & Crawl

Structured Platform Data

novada_scrape ships 13 built-in platform scrapers with structured data extraction (and the wider Novada Scraper API is reachable via novada_scraper_submit). Returns clean tabular records, not raw HTML.

Full platform list: call novada_discover or read the novada://scraper-platforms MCP resource.

Proxy Network

Route your own HTTP requests through Novada's proxy infrastructure. 100M+ IPs across 195 countries.

Each proxy tool returns connection credentials in url, env, or curl format. Params: country (ISO 2-letter), city (optional), session_id (for sticky sessions).

Browser Automation

For raw rendered HTML from protected pages, use novada_extract({url, format: "html", render: "render"}) (or render: "browser"). Sessions reuse the same page via session_id (cookies, login, page context) on the local/long-lived server; on the hosted serverless endpoint treat each call as one-shot.

Use Cases

AI Agent Research & RAG Pipelines

novada_research({question: "What are the latest developments in quantum computing?", depth: "comprehensive"})

Returns a cited multi-source report. Feed directly into RAG vector stores or use as context for agent reasoning.

E-Commerce Price Monitoring

novada_monitor({url: "https://amazon.com/dp/B0XXXXXX", fields: ["price", "availability"]})

First call records baseline. Call again later — returns field-level diffs with percentage change (e.g., price: $999 → $899, ↓10%).

Competitive Intelligence

novada_scrape({platform: "amazon.com", operation: "amazon_product_keywords", params: {keyword: "wireless earbuds"}, limit: 20})

Get structured product data (price, rating, reviews, BSR) for competitive analysis across the built-in platform scrapers and the wider Novada Scraper API.

Lead Generation

novada_scrape({platform: "linkedin.com", operation: "linkedin_company_information_url", params: {url: "https://linkedin.com/company/..."}, limit: 1})

Extract company info, employee count, and industry data from LinkedIn company pages.

Content Extraction for LLM Training

novada_crawl({url: "https://docs.example.com", max_pages: 20, select_paths: ["/docs/**"]})

Crawl documentation sites and extract clean markdown for fine-tuning datasets or knowledge bases.

AI Brand Monitoring

novada_ai_monitor({brand: "YourProduct", models: ["chatgpt", "perplexity", "claude"]})

Searches AI-company public web domains (openai.com, perplexity.ai, anthropic.com…) for public mentions of your brand: sentiment signals, claims, competitor co-mentions, source URLs. Note: this searches indexed public pages — it does NOT query the live AI models' responses; a brand with few indexed pages shows low/zero mentions.

Geo-Targeted Data Collection

novada_proxy_residential({country: "DE", city: "berlin", format: "curl"})

Get proxy credentials for any of 195 countries. Use with your own HTTP client for geo-specific content access.

Why Novada — Honest Comparison

Where competitors still lead: Firecrawl and Tavily return higher raw-character counts and lower P50 latency on easy pages (see benchmark/results/latest-summary.json); BrightData ships more structured scrapers (437 vs Novada's 13 built-in, with more reachable through the Scraper API). Novada trades raw volume for clean main-content extraction and a 4–5× lower cost, and is closing the latency gap (success rate climbed 70%→91% across the last two benchmark runs). Some Scraper API platforms need separate activation on your key.

Anti-Bot Support

Novada automatically handles these anti-bot systems via its escalation chain:

30+ domains are pre-tagged in the hard target registry — these skip static fetch entirely and go straight to the right method.

Staying on the latest version

Always use npx -y novada-mcp@latest — never npx novada-mcp (bare) and never npm install -g novada-mcp.

Why: A globally-installed novada-mcp shadows npx, so every subsequent npx novada-mcp silently runs the global binary regardless of what npm has published. You could be running a build that is multiple versions behind without any warning.

If you see an old version:

# 1. Check what you are running vs what npm has
npx novada-mcp@latest --version
npm view novada-mcp version

# 2. Remove any stale global install
npm uninstall -g novada-mcp

# 3. Clear the npx cache
npm cache verify

# 4. Confirm the correct version now runs
npx -y novada-mcp@latest --version

Links

• Docs + API key: novada.com
• npm: npmjs.com/package/novada-mcp
• Hosted MCP endpoint (beta): https://mcp.novada.com/mcp?token=YOUR_KEY
• GitHub: github.com/NovadaLabs/novada-mcp
• Issues: github.com/NovadaLabs/novada-mcp/issues
• Benchmarks: benchmark/results/latest-summary.json — auditable per-scenario numbers vs Firecrawl & Tavily
• Tool details: call novada_discover or novada_account from any MCP client

License

MIT

Quick Start

1. Get a key at novada.com

2. Add to your MCP client:

Claude Code:

claude mcp add novada -e NOVADA_API_KEY=your_key -- npx -y novada-mcp@latest

Claude Desktop / Cursor / VS Code / Windsurf:

{
  "mcpServers": {
    "novada": {
      "command": "npx",
      "args": ["-y", "novada-mcp@latest"],
      "env": { "NOVADA_API_KEY": "your_key" }
    }
  }
}

Hosted (no install, beta): point any Streamable-HTTP MCP client at https://mcp.novada.com/mcp?token=YOUR_KEY. A 15-tool subset, zero local setup. novada_browser runs on the hosted endpoint (via cloud CDP, one-shot per call); only novada_browser_flow (persistent multi-call sessions) and the disk-writing novada_site_copy are local-only.

3. Try it:

novada_search({query: "Claude MCP tutorials", num: 5})
novada_research({question: "How do MCP servers work?", depth: "deep"})
novada_extract({url: "https://news.ycombinator.com", format: "markdown"})
novada_monitor({url: "https://amazon.com/dp/B09...", fields: ["price", "availability"]})

Configuration