It has a JSON mode that you need to enable in settings and then you can create a simple python script to interact with it or have the agent use `curl` and `jq` to interact with it.
I am also interested in what a full local AI stack with web search and other tools looks like. As far as I can tell, SearX does not embed an MCP server, so it can't be directly called from llama-server for example. Open WebUI does have an integration for SearX and other providers, but the results I obtained weren't particularly impressive.
I use Searxng through Onyx, both as regular search and Onyx's Deep Research mode. I also have https://github.com/ihor-sokoliuk/mcp-searxng to add search to coding agents. Haven't really had many issues with it.
i have a friend with a 4080 that is wanting to experiment with local models and those cards should be similar enough. can you give any more detail about your setup? ty!
It is really great to use. As the poster above mentioned, my setup with Sear is the following, all through `llama.cpp`, which has a built-in webui with an MCP client:
* SearXNG in Docker — enable the JSON API (`search.formats: [html, json]`; off by default).
* `llama-server` with `--webui-mcp-proxy`, then add the server in the webui.
Some gotchas:
* `searxng-mcp` forgets to declare its own dep → `--with fastmcp`.
* Endpoint is `/mcp`, not the `/searxng-mcp/mcp` the docs claim.
* `--webui-mcp-proxy` only enables the CORS proxy; each MCP server entry still needs its "Use llama-server proxy" checkbox ticked, or the browser fetches direct and CORS-fails.
* Terminal clients (OpenCode etc.) skip the proxy — point them straight at `:8100/mcp`.
A couple interesting tidbits:
* There are temporal issues with search-related tool calls. The model trips out. 2026 results read to it a "future-dated hallucination" because it doesn't know the date. There's an additional `--tools get_datetime` function that will allow it to ground via the real date.
* Snippets-only is enough for most "what's current" questions and keeps context tiny.
I'm curious what setups folks use to provide this functionality.
Since the quantized 24B parameter Gemma model came out, I've had good luck with tool calling on a 4070 Ti Super.
Successful tool calling is what finally made the local experience useful.
I should note this is for the general and not coding specific context.