I've been using the LoFreq* caller for exactly this. It is designed to find variants with very low frequency, so is well suited for this type of analysis.
LoFreq* (i.e. LoFreq version 2) is a fast and sensitive variant-caller for inferring SNVs and indels from next-generation sequencing data. It makes full use of base-call qualities and other sources of errors inherent in sequencing (e.g. mapping or base/indel alignment uncertainty), which are usually ignored by other methods or only used for filtering.
LoFreq* is very sensitive; most notably, it is able to predict variants below the average base-call quality (i.e. sequencing error rate). Each variant call is assigned a p-value which allows for rigorous false positive control. Even though it uses no approximations or heuristics, it is very efficient due to several runtime optimizations and also provides a (pseudo-)parallel implementation.
(In my experience it is not fast at all, but that's what they claim).
I have used it many times to call somatic variants using the germline caller (and not its somatic one which requires a tumor/normal pair). Because the tool is so sensitive, its germline caller actually deals quite well with somatic samples.
You can find instructions for its use here.