We study a model of constraint satisfaction problems geared towards instances with few variables but with domain of unbounded size (udCSP). Our model is inspired by recent work on FPT algorithms for MinCSP where frequently both upper and lower bounds on the parameterized complexity of a problem correspond to $k$-variable udCSPs; e.g., the FPT algorithms for Boolean MinCSP (Kim et al., SODA 2023) and Directed Multicut with three cut requests (Hatzel et al., SODA 2023) both reduce to k-variable udCSPs, and the canonical W[1]-hardness construction in the area, Paired Min Cut by Marx and Razgon (IPL 2009), is effectively a k-variable udCSP. The udCSP framework represents constraints with unbounded domains via a collection $\mathcal{M}$ of unary maps into a finite-domain base language $Γ$. We develop an algebraic theory for studying the complexity of udCSP$(Γ,\mathcal{M})$ with a Galois connection based on partial multifunctions. We study three types of maps: unrestricted, one-hot, and monotone. For unrestricted maps, the problem is W[1]-hard for all but trivial cases, and for one-hot maps, the characterization coincides with Marx' FPT dichotomy for Boolean Weighted CSPs (Computational Complexity 2005). For the case of monotone maps Mo, we show that the complexity depends on restricted identifies we call ordered polymorphisms; we identify the "connector" polymorphism as the likely FPT boundary. We show that its absence implies that udCSP($Γ$,Mo) defines all permutations, and the problem is W[1]-hard; while its presence for a binary language implies bounded twin-width, and the problem is FPT (Twin-Width IV; Bonnet et al., JACM 2024). For non-binary languages, where twin-width does not apply, the polymorphism coincides with a notion of bounded projected grid-rank; however, we leave the FPT question for this case open.