Hi,
The physical I2C lines are open-drain. This means that the Pi pulls them up to 3.3 volt and the MCP pulls them low when communicating. So 5V from MCP never reaches the pi and the communication is safe, even then there is a voltage difference between host and slave. So although running the MCP at 3.3V seems to solve marcolastri’s problem, it should not matter…
I run several Pi’s with a MCP23017 powering the MCP at 5 volt, even driving a (5v) LCD-display from the MCP (Adafruit-like I2C CharLCDDisplay). On this display I do NOT use caps on the switches and it always works flawless. So could the debouncingcaps be the culprit?
Cheers,
Kees