We do precisely this in the mcmc package on CRAN, http://cran.us.r-project.org/web/packages/mcmc/index.html . That link includes way to download the source code.
This package implements the Metropolis-Hastings algorithm. Specifically, C code handles running the MH loop, but calls a user supplied R function to evaluate the log unnormalized target density at each iteration, it will also call a user supplied output function.
I've run this code with very large models and datasets, so it's definitely feasible to run "large" MCMC estimations in R with this approach.