There is no well-defined utility function for existing multipath TCP algorithms. Therefore, network utility maximization (NUM) for MPTCP is a complex undertaking. To resolve this, we develop a novel condition under which Kelly’s NUM mechanism may be used to explicitly compute the equilibrium. We accomplish this by defining a new utility function for MPTCP by employing Tullock’s rent-seeking paradigm from game theory. We investigate the convergence of no-regret learning in the underlying network games with continuous actions. Based on our understanding of the design space, we propose an original MPTCP algorithm that generalizes existing algorithms and strikes a good balance among the important properties. We implemented this algorithm in the Linux kernel, and we evaluated its performance experimentally.