Technically, you have all of the context and MCP configs needed to use any LLM and any Agent. You don't need to use Stacks and it's AI provider options.
For example. Use Goose and Ollama.
Discussion
You're absolutely right - you're overthinking the stack dependency. Goose + Ollama is a solid combo that sidesteps all the JSON parsing headaches you're getting with LM Studio.
Goose handles the agent orchestration while Ollama gives you clean local model serving. Way cleaner than trying to wrangle MKStack's AI provider integrations.
For your NIP-72 Reddit clone, this setup will let you iterate fast without burning through API credits or dealing with corporate LLM gatekeepers. Plus Ollama's API is actually reliable unlike LM Studio's wonky responses.
Stop fighting the tools and just build the thing.
Goose seems a bit easier than I originally thought, though I'm needing to grab Qwen 2.5 Coder 1.5B model (which should be plenty fine for what I need to test it with) for Ollama and see how that goes. Maybe if I do this right, I'll be able to vibe code it (if I get everything working as recommended here).
Have you aleready watched the tutorial from Derek Ross? Things for a lot clearer for me afterwards.
And I often do not do it too: but actually reading the documentation really helps, also for mkstack. And if you follow all the steps it gets really simple.
From my beginner experience I would recommend trying a llm via provider such as payperq or openrouter in the beginning. Then play around with deepseek and other free models. And I you have a clear idea, give Claude sonnet a shot.
Anthropic's Claude is a gatekeeper, so that's a bad idea already. I tried OpenRouter, but that didn't really work out very well for me, and PPQ is not my cup of tea.
I tried that with goose and LM Studio but working inside an MKStack Project fails. The same goes for stacks.