The MVar API is a bit inconsistent with the IORef one, and it catches me out every time. We have: modifyIORef :: IORef a -> (a->a) -> IO () which takes a nice pure modification function, and yet modifyMVar has a slightly weird Kleisli type: modifyMVar :: MVar a -> (a -> IO (a, b)) -> IO b which not only moves the modification into the IO monad but also permits returning an auxiliary value. One has only to browse the source of Chan to see that this function is useful, but I do think it's a surprising default. Then there is the _ version. So, we currently have: modifyMVar :: MVar a -> (a -> IO (a, b)) -> IO b modifyMVar_ :: MVar a -> (a -> IO a) -> IO () I suggest that the following would be more consistent with the normal conventions for _ and M: modifyMVar :: MVar a -> (a -> a) -> IO () modifyMVarM :: MVar a -> (a -> IO (a, b)) -> IO b modifyMVarM_ :: MVar a -> (a -> IO a) -> IO () ...however, that would break currently working code. If that seems to painful, then please at least supply some form of MVar a -> (a -> a) -> IO (), under whatever name seems suitable. Jules PS : Dear moderator : I subscribed and reposted, you may have a copy of this in your moderation queue too.