How I solved Cursor's hallucination problems with Supabase using 2 MCP Protocols

Discussion
1

Hi fellow cursor users!

I wanted to share some interesting findings from my recent experiments with Cursor AI and Supabase integration. Like many of you, I’ve been frustrated with Cursor occasionally hallucinating when generating React components, especially when dealing with complex database schemas.

I’ve been working on a solution using the Model Context Protocol (MCP) that has dramatically improved the accuracy and reliability of AI-generated code. The key breakthrough came from creating a dedicated MCP server that acts as a “truth source” for Supabase schema information.

What’s Different About This Approach

  • Real-time Schema Information: Instead of relying on Cursor to interpret database structures from code snippets, the MCP server provides real-time, accurate schema information.
  • Understanding Complex Relationships: The server comprehends complex relationships, RLS policies, and type constraints, preventing common AI hallucinations.
  • Accurate React Components: Generated React components are always in sync with the actual database structure.
  • Elimination of Manual Fixes: No more back-and-forth fixing incorrect type assumptions or mismatched field names.

Some Unexpected Benefits I Discovered

  • Precise TypeScript Types: The AI generates much more precise TypeScript types since it has direct access to the schema.
  • Automatic RLS Policy Integration: RLS policies are automatically considered when generating data fetching logic.
  • Maintained Foreign Key Relationships: Foreign key relationships are properly maintained in forms and data displays.
  • Immediate Schema Updates: Schema changes are immediately reflected without needing to update context files.

I’ve also developed a dynamic context system using gemini-2.0-pro-exp that automatically updates as your codebase context and instructions evolve, which has been a game-changer for larger projects. The AI seems to understand the codebase much better than with static @ Codebase references.

Questions for the Community

  • Has anyone else experimented with MCP for improving AI accuracy?
  • What are your biggest pain points with Cursor generating database-connected components?
  • Would you be interested in seeing this released as an open-source tool?

I’m particularly curious about your experiences with Cursor hallucinations in database-heavy applications and how you’ve addressed them.

2 Likes
2

Can you share the server?

2 Likes
3

“I have the magic wand my dudes!!”

4

Ok sure MCP is from now on going to be access for additional info thats not built in. Would you mind sharing a bit more about your approach to help the community?
Example, what the 2 protocols are doing as one for supabase would be enough.

We have tried several approaches regarding codebase without MCP as it wasnt yet practical, but mostly it evolves around reducing tokens, creating docu,…

Depending on project, the DB structure we have is already documented in language specific files and updated from there, so no realtime part required.

5

Yes OS pls!