Kindly Web Search Mcp Server

The honest selling point here is not search, it is the extraction layer. Raw SERP text is context poison: it burns tokens, adds markup noise, and measurably degrades reasoning. Shelpuk's own write-up on the problem claims agents fall to 20-30% accuracy on commercial codebases, and Kindly exists to fix that failure mode by returning clean structured JSON instead of a wall of links. We would reach for this when an agent needs current facts mid-task: checking a library's latest API, reading changelog entries, confirming a version number before generating code. In practice the caching and rate-limit controls matter more than the search provider choice, because a loop-happy agent can hammer an API key into a spend cap fast. Where it bites is setup. This is a self-hosted MCP server, not a sign-up-and-go service. You configure it via CLI and config files, wire in your own Serper or Tavily key (or stand up SearXNG), and manage the JavaScript rendering path yourself. GitHub Copilot users who want zero infrastructure work will find that friction real. The closest alternative in this space is wiring a hosted search API straight into your agent. That is simpler up front and gives you no caching, no domain/date/type filtering, and no self-hosted privacy option. Kindly trades a Saturday of setup for control over where your queries go and what shape they come back in. For European teams with data residency requirements, the self-hosted SearXNG backend is the feature that actually matters: no per-query cost and no third-party search vendor seeing your agent's questions. For everyone else, it is a nice fallback rather than the headline. Choose it if you already run an MCP workflow and want live web access inside it. Pass if you need a graphical product, an SLA, or compliance