The cache is the mechanism that makes the five-tool design pay off: execute costs credits once, and then get_page, query_cache and export_data all run server-side against the 7-day cache without re-charging — so narrowing and exporting a large result is cheap. The discover-before-execute pattern is what lets one small tool set cover dozens of sources: the agent reads the endpoint's parameters from discover rather than needing a bespoke tool per site.
Anysite's official server, exposing dozens of data sources through five universal tools plus an optional set of CRM write-back tools. It runs as a hosted Streamable HTTP server with OAuth, with an npm stdio bridge for clients that only speak stdio.
- discover lists the endpoints in a category for a source, with their parameters and LLM hints — always called before execute, so the agent learns the parameter names on its own rather than guessing
- execute fetches data from a source's endpoint and returns the first 10 items plus a cache_key
- get_page paginates the cached items without re-fetching (limit up to 50), so one execute call feeds unlimited paging without re-paying credits
- query_cache filters, sorts and aggregates the cached items server-side by conditions, sort_by, aggregate and group_by, narrowing a large result set without pulling it all back
- export_data dumps the full dataset to json, jsonl or csv and returns a download URL
- Sources surfaced through discover include LinkedIn, Instagram, Twitter/X, Reddit, YouTube, SEC EDGAR, Y Combinator, Crunchbase and DuckDuckGo, plus a universal web parser and AI parsers for GitHub, Amazon, Google Maps, G2 and more — new endpoints appear automatically as they ship
- An optional set of eight crm_* tools for HubSpot-and-friends read/write — connections, schema, upsert of contacts and companies, query and undo — visible only when the CRM integration is enabled on your account
- Fetched data is cached server-side for 7 days, so paging, querying and exporting off one execute call cost no additional API credits
An Anysite account from anysite.io on a paid MCP plan. Connect directly to the hosted server at https://mcp.anysite.io/mcp over Streamable HTTP with OAuth — if your client speaks HTTP you do not need the npm package. For stdio-only clients, the @anysiteio/mcp bridge needs an ANYSITE_ACCESS_TOKEN, which can live in ~/.anysite-mcp.env. In Claude Desktop, set tool access to 'Tools already loaded', or the generic names discover and execute will not be reliably matched.
One command plus a key — npm install -g @anysiteio/mcp, then supply credentials