Since mesh colliders dont update in respect to the current hand pose this would be very taxing on performance to create at runtime. What id recomend is iterating through each finger joint after they have been initialized and attaching a capsule collider with appropriate dimensions. This will give you very accurate collision which conforms to any pose with very little overhead. Additionally you can set the colliders as triggers if desired while a mesh collider can only be convex in trigger mode which would negate the accuracy you want
I've been looking into this too. That rbgschemes post above is great. But worth noting it doesn't work with the current unity integration... the principles are probably the same however. I've yet to get it working but will follow up more if I can
