Would you use a local Perplexity-style setup like Vane with Ollama and SearXNG for private research, or is the convenience of hosted AI search still worth the tradeoff?
I have a MacBook Pro, 40-core GPU, 48 GB RAM, M3 Max chip, and I'm running qwen3.6-35b-a3b as MLX 4bit.
I have the Hermes Agent from Nous Research and I switch between the terminal and the Hermes desktop app.
Now I want to set up web search. My research suggested that SearXNG and Firecrawl would be a good idea.
My question for you: since I already have a UI with the Hermes desktop app, should I build the search myself with SearXNG and Firecrawl, or does that get too complicated and Vane is the better choice? I want to search 100% locally and anonymously, so I'm also thinking about a VPN. Security is an important aspect too.
I have time to try everything out. The stack is meant to be built up for the long term. What's your advice for me?
Hi Bianca, great question, and your instinct is right: for a long-term local stack, I’d keep Hermes as your main UI and build the search layer underneath it first. Setting up Vane may just duplicate a lot of functionality you already have in your current setup.
My suggested path:
1. Start with local SearXNG for simple JSON-format search.
2. Add Firecrawl only if you need cleaner page extraction, crawling, or JavaScript-heavy page handling.
4. A VPN is easiest to implement if you run it at the macOS/network level so outbound web requests from SearXNG and Firecrawl go through the VPN, while Hermes and MLX handle the model locally. Tor is more anonymous, but often more fragile for search because of CAPTCHAs and blocks.
You can definitely test Vane separately just to try it out, but it probably won't give you sufficient integration with your current agent setup.
So my advice: build the modular SearXNG + Firecrawl route in stages, and consider Vane as a standalone search application.
We haven’t personally tested this exact Hermes + MLX + SearXNG + Firecrawl setup (yet) but it’s definitely the kind of workflow we may turn into a full article once we do.
Would you use a local Perplexity-style setup like Vane with Ollama and SearXNG for private research, or is the convenience of hosted AI search still worth the tradeoff?
So glad I found you!
I have a MacBook Pro, 40-core GPU, 48 GB RAM, M3 Max chip, and I'm running qwen3.6-35b-a3b as MLX 4bit.
I have the Hermes Agent from Nous Research and I switch between the terminal and the Hermes desktop app.
Now I want to set up web search. My research suggested that SearXNG and Firecrawl would be a good idea.
My question for you: since I already have a UI with the Hermes desktop app, should I build the search myself with SearXNG and Firecrawl, or does that get too complicated and Vane is the better choice? I want to search 100% locally and anonymously, so I'm also thinking about a VPN. Security is an important aspect too.
I have time to try everything out. The stack is meant to be built up for the long term. What's your advice for me?
Hi Bianca, great question, and your instinct is right: for a long-term local stack, I’d keep Hermes as your main UI and build the search layer underneath it first. Setting up Vane may just duplicate a lot of functionality you already have in your current setup.
My suggested path:
1. Start with local SearXNG for simple JSON-format search.
2. Add Firecrawl only if you need cleaner page extraction, crawling, or JavaScript-heavy page handling.
3. Keep everything bound to localhost at first.
4. A VPN is easiest to implement if you run it at the macOS/network level so outbound web requests from SearXNG and Firecrawl go through the VPN, while Hermes and MLX handle the model locally. Tor is more anonymous, but often more fragile for search because of CAPTCHAs and blocks.
You can definitely test Vane separately just to try it out, but it probably won't give you sufficient integration with your current agent setup.
So my advice: build the modular SearXNG + Firecrawl route in stages, and consider Vane as a standalone search application.
We haven’t personally tested this exact Hermes + MLX + SearXNG + Firecrawl setup (yet) but it’s definitely the kind of workflow we may turn into a full article once we do.
Good luck, and happy hacking!