As the title asks, I'm wondering if one can generally squeeze a "covariant power object monad" out of a topos (following the usual example in $\mathcal{Set}$ with functor part the direct image powerset functor), or if there's a nice counterexample showing why not.
I am guessing that the answer to this is negative, since I only see this monad discussed in the case of $\mathcal{Set}$ so I suspect it depends on more special properties of that category; in which case I'm curious what special properties make it work.
Edit: In particular, I understand the usual construction of the functor part, but have myself been unable to verify that the obvious candidates for the unit and multiplication actually work as such.
In fact, the answer is yes. The algebras for the monad are, of course, the internal complete join semilattices.
First things first: the unit $\eta_X : X \to P X$ is given by the transpose of the morphism $X \times X \to \Omega$ that classifies the diagonal $\Delta : X \to X \times X$. To see that it is a natural transformation, it is easiest to use internal logic: given $f : X \to Y$, we want to show $$x : X \vdash \exists_f (\eta_X (x)) = \eta_Y (f (x))$$ which by "extensionality" is equivalent to $$x : X, y : Y \vdash y \in \exists_f (\eta_X (x)) \leftrightarrow y \in \eta_Y (f (x))$$ which by definition of $\exists_f$ is equivalent to $$x : X, y : Y \vdash (\exists x' : X . x' \in \eta_X (x) \land f (x') = y) \leftrightarrow y \in \eta_Y (f (x))$$ which by definition of $\eta$ is equivalent to $$x : X, y : Y \vdash (\exists x' : X . x' = x \land f (x') = y) \leftrightarrow y = f (x)$$ which is obviously a tautology.
The multiplication $\mu_X : P P X \to P X$ has a simple description in internal logic: $$\mu_X (t) = \left\{ x : X \mid \exists s : P X . s \in t \land x \in s \right\}$$ If you unfold this, it amounts to saying that $\mu_X$ is the transpose of the morphism $P P X \times X \to \Omega$ that classifies the image of the composite $$R \rightarrowtail P P X \times P X \times X \to P P X \times X$$ where the second arrow is the obvious projection and the first arrow is defined by the following pullback diagram, $$\require{AMScd} \begin{CD} R @>>> [\ni] \times [\ni] \\ @VVV @VVV \\ P P X \times P X \times X @>>{\mathrm{id} \times \Delta \times \mathrm{id}}> P P X \times P X \times P X \times X \end{CD}$$ where $[\ni] \rightarrowtail P P X \times P X$ and $[\ni] \rightarrowtail P X \times X$ are the universal binary relations. Of course, it would be a nightmare to verify that $\mu_X$ defined this way is natural in $X$, so it's better to stick with the description in internal logic. I leave the verification to you – naturality amounts to saying that $\exists$ commutes with $\exists$.
It remains to be shown that $\eta$ and $\mu$ satisfy the monad axioms. It should go without saying that the best way to proceed is to use internal logic.