The hard parts of scraping — rendering, geography, blocking — sit on Decodo's side, so the conversation deals in results. Two knobs are worth setting deliberately: the toolsets parameter, because registering forty-odd tools costs context on every turn, and the token limit, which keeps a large page from filling the window.
Decodo's own scraping server, organised into five toolsets — general web, search engines, ecommerce, social media and AI answer engines — over its proxy and rendering network.
- Any URL scraped and returned as Markdown, or captured as a PNG screenshot
- Search results parsed rather than raw: Google search, ads, Lens, AI Mode and Travel Hotels, plus Bing
- Ecommerce coverage with per-page structure: Amazon search, product, pricing, sellers and bestsellers; Walmart and Target search and product; TikTok Shop search, product and URL
- Social platforms as structured data: Reddit posts, subreddits and user profiles; TikTok posts with engagement, caption and hashtags; YouTube metadata, channel videos, subtitles and search
- AI answer engines queried directly — ChatGPT, Perplexity and Google AI Mode
- Request behaviour inferred from your prompt: headless rendering, the country the request comes from, its locale, and a token limit that truncates the response for a small context window
- Toolsets you can narrow to, so only the categories you need are registered at all
A Decodo account and its basic auth token from the dashboard. Remote is the quickest path: point the client at https://mcp.decodo.com/mcp with an Authorization header, adding a toolsets query parameter to limit what is exposed. To run locally, use npm as @decodo/mcp-server over stdio with SCRAPER_API_TOKEN, and TOOLSETS to select categories. With no toolsets named, everything is registered.
One command plus a key — npx -y mcp-remote https://mcp.decodo.com/mcp, then supply credentials