oMLX and Pi: Search MCP Configuration 2026

oMLX and Pi are popular agent harnesses for local LLMs in 2026. Both support MCP for adding external tools like web search. Configure Tavily MCP (1,000 free/month, community maintained) or Scavio MCP (hosted at mcp.scavio.dev/mcp, multi-platform) to give your local Qwen, Gemma, or Llama model internet access.

oMLX MCP configuration

oMLX 0.39+ supports standard MCP server definitions. Add a search server to your oMLX config file:

{
  "mcpServers": {
    "scavio-search": {
      "type": "url",
      "url": "https://mcp.scavio.dev/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_SCAVIO_API_KEY"
      }
    }
  }
}

Pi extension: zero-config option

Pi supports extensions that add capabilities without MCP configuration. The pi-web-access extension from nicobailon provides web search using Exa as the default backend:

pi install npm:pi-web-access

This works out of the box with no API key configuration. For more control over the search backend, use an MCP tool instead.

Tavily MCP for oMLX

Reddit users confirmed Tavily MCP works in both oMLX and Pi. The community-maintained server connects to Tavily's 1,000 free searches/month tier:

{
  "mcpServers": {
    "tavily": {
      "command": "npx",
      "args": ["-y", "tavily-mcp-server"],
      "env": {
        "TAVILY_API_KEY": "tvly-your-key"
      }
    }
  }
}

Which search backend for local LLMs

• Tavily: best for general Q&A grounding. Summarized results work well with smaller models that struggle to parse raw search data.
• Scavio: best for structured data needs. Returns typed JSON that local models can extract specific fields from (prices, dates, URLs).
• Exa: best for semantic search. Finds conceptually similar content rather than keyword matches.
• SearXNG: free self-hosted option. Requires running another container alongside your model server.

Model-specific notes

From the LocalLLaMA community: Gemma 4 31B handles tool calling better than Qwen 3.6 27B for search-augmented workflows. Qwen 3.6 35B A3B is a good budget option. For models under 14B parameters, Tavily's summarized results produce better outputs than raw search JSON because smaller models struggle to extract relevant information from verbose structured data.