If there were some consonant than only appeared in clusters, the standard analytic assumption would be to reduce it to some consonant that does not appear in clusters, thus if [γ] only appeared in clusters and [h] never did, then we would simply assume that [γ] comes from /h/. To show that such an analysis is impossible, you'd have to have a language that allows pretty much any possible cluster, which is extremely rare. That said: glottal stop only appears before nasals in North Saami. Which leads to another path of re-analysis, where "glottal stop" is interpreted as the surface realization of glottalization on nasals.
Although the question is perfectly sensible, I think there are so many ways to wiggle that it is effectively impossible to answer.
Following up given OPs question... I'd be happy with rewording to eliminate wiggle-room, but I don't know what that wording would be. I don't think the problem is ambiguity in the question, it's the looseness of our underlying theories. Reduction in underlying inventory has long been a desideratum of analysis, hence we don't posit both aspirated and unaspirated voiceless stops in the English underlying inventory. This means that you are virtually compelled to analyze out of existence putative cases of segments with defective distribution. What principle could we adopt that would prevent this practice (and should we)?
If the question is stated in terms of phonetic outputs, then the answer is "there are plenty", such as English [ɬ] which only appears after [s,f] and aspirated stops; or [ʈ] (in some dialects, phonetic value approximate) which only appears before /ɹ/. So to favor a positive answer to the question, we would need to limit abstract analysis; to favor a negative answer, we would encourage (require) such analysis.