I recommend taking 'mkFooC' out of the typeclass.  It keeps you from being able to (easily) construct a 'FooC' from dynamic data, e.g.:

    mkFoo :: Host -> Port -> IO MyFoo

After this change, the typeclass approach and the data constructor approach are nearly equivalent, except:

 * With the typeclass approach, the compiler passes the dictionary implicitly, which can be more convenient to use (e.g. `readFooC a` instead of `readFooC (getFoo a)`).

 * With the typeclass approach, you have to define a Foo type to contain the environment needed for Foo methods.  With the record approach, you can just construct and use a FooT record directly.

Either way, don't forget about simple encapsulation:

    data LineDevice -- abstract

    -- Some LineDevice constructors for common tasks

    stdio :: LineDevice

    openFile :: FilePath -> IO LineDevice

    connectTo :: HostName -> PortId -> IO LineDevice

    getLine :: LineDevice -> Int -> IO ByteString

    putLine :: LineDevice -> ByteString -> IO ()

This interface is very easy to understand.  If you want to let users make their own LineDevice objects, you can still provide an "internal" module with something like this:

    data Driver = Driver

        { getLine :: Int -> IO ByteString

        , putLine :: ByteString -> IO ()
        }

    newLineDevice :: Driver -> IO LineDevice

Hope this helps,

-Joey