Define the midpoint $m(u,v)$ and swivel $\phi_{u,v}(w)$ by
$$ m(u,v)=\frac{u+v}{2}, \quad \phi_{u,v}(w)=\frac{m(u,v)w-uv}{w-m(u,v)}. $$
It is easy to verify $\phi_{u,v}$ sends $u,v,m(u,v)$ to $u,v,\infty$ respectively. Mobius transformations act sharply $3$-transitively (the orientation-preserving ones, to be precise), so this uniquely characterizes $\phi_{u,v}$.
Indeed, from the classification of Mobius transformations, we can interpret $\phi_{u,v}$ as a $180^\circ$ "rotation" around poles $u$ and $v$ in bipolar coordinates. This doesn't require $u,v$ to be finite: in the case of $\phi_{\infty,v}(w)=2v-w$, "bipolar" coordinates are just polar coordinates, and this is a typical $180^\circ$ rotation of $w$ around $v$. Since the Apollonian circles of bipolar coordinates are determined by the two poles, and Mobius transformations stabilize the set of (generalized) circles, we can say that $g\circ\phi_{u,v}\circ g^{-1}=\phi_{gu,gv}$ for any other Mobius transformation $g$, by symmetry.
Then we can construct the "duality" map $\Phi(a,b,c)=\bigl(\phi_{b,c}(a),\phi_{c,a}(b),\phi_{a,b}(c)\bigr)$. It is defined for any $a,b,c$ distinct (note collinearity is allowed). If we extend the action of Mobius transformations to triples, we then have $g\circ\Phi\circ g^{-1}=\Phi$ for any transformation $g$. Since transformations act triply transitively, to verify $\Phi^2=I$ it suffices to show $\Phi^2$ fixes a particular triple, which we can choose to be the equilateral triangle $(a,b,c)=(1,\omega,\omega^2)$ (where $\omega$ is a primitive cube root of unity); it is easy to verify $\Phi(1,\omega,\omega^2)=(-1,-\omega,-\omega^2)$, and the result follows.
While I've noted a geometric interpretation of $\phi_{u,v}(w)$, I fail to see one for $\Phi(a,b,c)$ at the moment. It seems to be some kind of "dual" triangle,