One search endpoint for agents that can't browse

What it does

AnySearch MCP Server exposes web search, vertical domain search, and full-page extraction as remote tools that MCP clients can call directly. It answers open-ended natural-language queries, routes structured requests through specialized domains like finance, academic, and code, and fetches HTML pages as Markdown truncated at 50,000 characters. The whole thing runs as a hosted service, so the client just points to an endpoint rather than managing its own scraping stack.

The interesting bit

The server implements the March 2025 MCP spec’s Streamable HTTP transport natively, which means newer clients such as OpenCode and Claude Desktop 2025.6+ can talk to it directly. Everyone else—Cursor, Windsurf, legacy Claude—has to route through a local stdio or SSE proxy, making this feel like a modern protocol with a compatibility shim for the installed base.

Key highlights

• Anonymous access works out of the box with reduced rate limits; an optional API key simply raises the ceiling
• batch_search fires up to five independent queries in parallel and isolates failures so one bad request doesn’t kill the batch
• Vertical search requires agents to call get_sub_domains first to retrieve valid routing keys and parameter schemas, enforcing structured inputs instead of guesswork
• extract converts full HTML pages to Markdown, though it is strictly HTML-only and capped at 50,000 characters
• Native Streamable HTTP support means no local proxy for modern clients; stdio and SSE users need third-party proxy tooling

Caveats

• Domain-specific searches are gated behind a mandatory get_sub_domains lookup, adding a round-trip agents must handle explicitly
• The extract tool only consumes HTML and truncates hard at 50,000 characters, ruling out large documents or JavaScript-heavy apps
• Clients stuck on stdio or SSE transports need to keep a proxy process running locally, which is extra moving parts

Verdict

Handy if you’re wiring up an MCP agent that needs live web data but you’d rather not maintain your own search or scraping pipeline. Less compelling if your client already has robust built-in browsing or you need to extract content beyond static HTML.