This piece of magic is the result of a DC servo loop that is wrapped around the input of the current source and the output of the Op-Amp. If the input moves toward the positive over a long period of time, that net DC drift is fed into the non-inverting input of the Op-Amp, which causes its output to go positive. This positive voltage will further drive the MOSFET into greater conduction, which will pull the output towards negative. On the other hand, if the output moves toward the negative over a long period of time, that net DC drift is fed into the same non-inverting input of the Op-Amp, which causes its output to go negative. This negative voltage will move the MOSFET into less conduction, which will move the output towards positive.
     The subtlety here is that the MOSFET is within the DC feedback loop, but outside the AC feedback loop of the Op-Amp. As the time constant of the RC network made up of the two 1 meg resistors in parallel and the .1 µf capacitor is so long that no music can fall into it, the Op-Amp presents a virtually constant DC voltage to the gate of the MOSFET. This steady voltage sets the amount of current that flows through the MOSFET; if the tube's idle current drifts over

Compliant Current Sources
     If by adding sufficient circuitry we can eliminate the coupling capacitor, then the added complexity might well be worth the effort. A DC servo loop can be wrapped around the current source, rather than the tube. This creates a compliant current source that tracks the tube's quiescent current and nulls a DC offset at the output. This circuit was covered in a GlassWare Tube circuit of the Month article, No Gain, No Pain and is quoted here:
   This trick consists of a Cathode Follower that is loaded at its output by a compliant current source, that is a current source that does not have a predetermined quiescent current. What is constant about it is that it strives to maintain a DC ground potential input regardless of the current flowing through it.
     Is it then really a current source? In AC terms, yes; in DC, no. It offers a very high impedance (roughly, 1 meg) to any AC signal it sees and in this respect it is identical to a typical active current source. In DC terms it works to adjust its quiescent current until its input is zeroed at ground potential (0 volts) over time (roughly, 3 Hz).