Novada MCP

Novada MCP

One MCP server for the entire live web. Search, extract, scrape, crawl, proxy, browser automation, and AI-powered research — behind a single hosted connection, or one local install if you'd rather run it yourself.

Quickstart · Get your key · How to choose a tool · Why Novada · Repository layout · Links

Quickstart

Novada is hosted-first — there's nothing to install. Point your client at the hosted URL and you're done.

Security note: the hosted endpoint never accepts your API key as a URL query parameter or path segment — only an Authorization: Bearer <key> header, or OAuth sign-in, is accepted. This keeps the key out of URLs, server logs, proxies, and browser history. Never share your key or paste it into a URL.

claude.ai (web)

1. Go to Settings → Connectors → Add custom connector.
2. Name it Novada.
3. Paste the bare URL and sign in with your Novada account when prompted (OAuth):

    https://mcp.novada.com/mcp

4. Click Add.

Claude Code

claude mcp add --transport http novada https://mcp.novada.com/mcp --header "Authorization: Bearer YOUR_KEY"

Full per-client install (Cursor/Windsurf/VS Code + local self-host) → npm-package/README.md.

Try it

novada_setup()                                          — validates your key, shows balance
novada_search({query: "Claude MCP tutorials"})          — web search
novada_extract({url: "https://example.com"})            — read any URL
novada_research({question: "how do MCP servers work?"}) — parallel multi-source research

Get your key

1. Sign up at novada.com — no credit card required.
2. Copy your API key from novada.com.
3. You start with $10 in free credits, capped at 1,000 calls/month. Check what's left any time with novada_account({section: "balance"}) or novada_setup().

How to choose a tool

These are the tools you reach for most:

Not sure? Call novada_discover and it returns the full catalog with each tool's status.

📚 Full reference — all 37 tools →

37 tools across 6 categories. Self-host (npx novada-mcp) exposes all 37. The hosted default surface (mcp.novada.com) exposes 28 — the same registry minus 9 tools that don't apply to a stateless serverless endpoint: novada_browser / novada_browser_flow (both need a persistent CDP WebSocket session a serverless isolate can't hold), novada_site_copy (writes files to disk), novada_ip_whitelist / novada_static_ip_mgmt / novada_capture_apikey (write-gated account ops), novada_session_stats / novada_search_feedback (per-process in-memory state), and novada_verify — it is core-derived, not a hand-curated subset. Call novada_discover on your connection to see exactly what's available on it.

Where does it run?

Full local/hosted breakdown, all 37 tools

Full per-tool reference (same column) → docs/TOOLS.md.

Why Novada

• 14 dedicated per-platform scrapers. novada_scrape_amazon, _google, _bing, _duckduckgo, _yandex, _youtube, _instagram, _facebook, _tiktok, _x, _walmart, _linkedin, _github, _perplexity — each exposes a closed, typed operation enum scoped to that platform instead of the generic novada_scrape's open platform+operation string pair, so an agent can't guess an invalid operation for the wrong platform.
• Honest tool surface. Every listed operation is verified-working — operations we can't currently deliver are marked and excluded, not left in to fail on you mid-task.
• Contract-tested self-report. Tool descriptions are tested against actual behavior, not just written and forgotten — what a tool claims to return is checked against what it actually returns.
• Drift-guarded tool registry. npm-package/src/tools/registry.ts is the single source of truth for the tool catalog; a test asserts the registered tools, the wired tools, and the novada_discover output can never diverge.
• confirm:true write-gate. Every mutating tool (proxy sub-account creation, IP whitelist changes, static IP purchases, capture-key resets) requires an explicit two-step confirmation — no silent writes.
• Callable onboarding. novada_discover and novada_setup are tools your agent can call itself to find the right tool or validate a key, without ever reading this README.

vs Firecrawl / Tavily (measured, not marketing):

• 4.9× more compact. An identical 5-result search query returned 2,761 bytes from Novada vs 13,421 bytes from Firecrawl — less agent token cost per answer.
• Keyless to start. npx -y novada-mcp cold-starts with no key; novada_setup walks you to one only when a tool needs it.
• Named anti-bot failures. On an unresolvable target, Novada's response names the exact blocker (e.g. anti_bot:kasada) instead of a generic error, so an agent can branch on it.

Repository layout

This is a monorepo with two artifacts:

• npm-package/ — the novada-mcp npm package. Local stdio MCP server (npx novada-mcp); source of truth for all tool logic.
• hosted-server/ — what runs at https://mcp.novada.com. An HTTP wrapper (auth, quota, rate-limit) on Vercel around the npm package's built output.

Full architecture map (entrances, dispatch core, where-does-X-live) → ARCHITECTURE.md; contributor routing → CONTRIBUTING.

Links

• Website: novada.com
• Get an API key: novada.com
• Sign up (free): novada.com
• npm: npmjs.com/package/novada-mcp
• GitHub: github.com/NovadaLabs/novada-mcp
• Issues: github.com/NovadaLabs/novada-mcp/issues

License

MIT