rm-rf-rm

I actually don't understand how supabase is a good backend choice for vibe coding, i've found there to be quite an unbelievable amount of friction to hook up supabase to AI agents: The MCP is ridiculously wasteful in tokens (pollutes every prompt to the tune of 6-7000 tokens). AI Agents dont know how to use the supabase CLI, constantly trying to run general `supabase db query "SELECT.."` - a command that doesn't exist. Worst, Even Opus 4.5 does not know that anon keys are legacy now. Im imagining its just the lovable, bolt grade apps whose engineering teams defaulted to making supabase their default backend? We're already in the Claude Code era, so I expect that wave to die out as time goes on.

I'd 100% use that as long as I can use local models (which are plenty capable for this use case)

This is 100% a major area that is yet very underserved and all options in the market kinda suck IMO (think Tableau or worse). That said, there's absolutely no way Im going to trust a SaaS with this type of priviledged access to my data. Even if I had an enterprise account with a rock solid contract

thanks for chiming in. I already do that and have reduced it to the basic 5 that are needed. Even so, thats >3k tokens which is an unacceptably large amount for just 1 MCP that is going in every single LLM call. Is there a reason for each tool requiring that many tokens? Surely a list tables tool - something that is a single command needs 600+ tokens.