AI tinkering - MCP

One of my apps, Wheel Flow, retrieves data from a backend in order to handle the day-to-day logic. For this specific feature, here's how it typically works:

1. User requests a ticker

2. App sends the request to the backend

3. Checks the database

4. If data's stale, triggers an edge function

5. Edge function calls an external API

6. Updates the database

7. Returns fresh data to the app, which stores it locally as needed

Recently I discovered a pricing update bug in certain stocks under certain conditions, it was pretty easy to understand why the bug happened, but rather than attempting a fix right away I though of using the opportunity to play with Model Context Protocol by hooking up Supabase with Cursor using MCP.

The experiment begins

I had several questions:

• Does it work? How?

• How far can read the data? Table schema? Data inside each table?

• Read-only access? Write?

• Can interact with edge functions?

• How flexible is this protocol, in general?

My first attempt following the documentation failed. It couldn't read the .cursor/mcp.json file to enable it, so I ran the command directly in Cursor Settings and voilá:

Some experiments to see what sort of access does it have through MCP:

Modifying tables and running SQL queries worked nicely, however, it cannot seem to be able to access edge functions, or perhaps there's a way that I don't know yet.

This meant I finally had to fix the bug manually and the MCP connection couldn’t help here, as it was a problem in the logic of one of the edge functions:

Final thoughts

Model Context Protocol isn't magic, at least not yet. In this specific case (the Supabase integration) it crushed it with database stuff, easily running SQL queries and modifying tables, but not being able to interact with edge functions makes it a bit lacking just yet.

While still a bit messy and clunky, it’s killer tool to add to the AI arsenal, the potential is real, and cannot wait to play with other MCPs.