When Claude’s Built-In Search Still Hallucinates

What it does

GrokSearch is an MCP server built on FastMCP that gives Claude Code and Cherry Studio a more aggressive web-search layer. It routes queries through Grok for AI search and Tavily for page scraping and site mapping, with Firecrawl as an automatic fallback when Tavily fails. The server exposes eight tools, including the ability to disable Claude Code’s own WebSearch and WebFetch by editing the project’s .claude/settings.json so your queries hit this pipeline instead.

The interesting bit

The README includes a side-by-side showing Claude Opus 4.6 ignoring its own built-in search and answering from internal knowledge rather than checking FastAPI’s official docs; only after being forced through this MCP does the model run multiple targeted searches. That willingness to kneecap the native tools via toggle_builtin_tools is a refreshingly direct piece of systems engineering.

Key highlights

• Dual-engine pipeline: Grok for AI search, Tavily for scraping and site maps, Firecrawl as automatic fallback
• Eight MCP tools including search planning, model switching, source retrieval by session ID, and config diagnostics
• toggle_builtin_tools edits Claude Code’s project settings to disable the built-in web tools and force routing here
• Auto-injects local time context when queries contain temporal keywords like “latest” or “today”
• Supports OpenAI-compatible Grok endpoints, including mirror stations

Caveats

• Tavily and Firecrawl are optional; without them web_fetch and web_map return configuration errors rather than degraded results
• Windows users are strongly urged to run under WSL rather than natively
• Enterprise or proxy environments may encounter SSL certificate errors that require passing --native-tls to uvx

Verdict

Useful if you run Claude Code and need a stickier way to make it verify facts against live documentation. Skip it if you don’t have access to an OpenAI-compatible Grok endpoint or you’re happy with your current search setup.