llama.cpp users debate next homelab stack

The sharpest answer here is that the jump from local model to useful assistant comes from tools, retrieval, and memory more than another orchestration layer. `llama.cpp` already gives you the engine: [llama.cpp](https://github.com/ggml-org/llama.cpp).

• –Open WebUI gets you a usable front end fast, but its own docs say agentic search works best with stronger models and that small local models can struggle: [Open WebUI agentic search](https://docs.openwebui.com/features/web-search/agentic-search/)
• –SearXNG plus a thin fetch endpoint is a simpler search layer than building bespoke retrieval glue, especially if you just want good web grounding.
• –LangGraph is the right hammer only if you truly need durable, stateful workflows; its docs position it as low-level orchestration with long-term memory and human-in-the-loop: [LangGraph docs](https://langchain-ai.github.io/langgraphjs/reference/modules/langgraph.html)
• –MCP is the clearest Claude-like upgrade because it standardizes tool exposure; the spec says models can invoke tools to query databases, call APIs, and run computations: [MCP tools spec](https://modelcontextprotocol.io/specification/2025-06-18/server/tools)
• –If the goal is local Claude, prioritize model quality, context length, and tool reliability before adding more layers.