CINT is the closest thing I've ever seen to a real C++ REPL. It's actually used quite extensively in the ROOT data analysis environment (REPL, reflection, UI signal/slot mechanism, object persistency, creating dynamically Python bindings, etc.).

As a REPL CINT actually works surprisingly well. Unfortunately it seems to have some problems interpreting certain "advanced" C++ stuff, such as templates and especially code that uses STL heavily. Another problem with it is that since you still need to manage memory manually and have direct access to pointers it's very easy to make CINT (or ROOT) session crash... :(

Btw, the ROOT developers are also working on a new LLVM based C++ interpreter called Cling: http://root.cern.ch/drupal/category/package-context/cling

It does. I use it to check quick c questions. It works kind of like gdb's 'p' command, but you can also execute c code by putting it in braces.

Here's an excerpt from my most recent cint repl session:

  cint> {long x = 84bfaad645e2e33f; double * p = &x;}
  Warning: Illegal numerical expression 84bfaad645e2e33f FILE:(tmpfile) LINE:1
  cint> {long x = 0x84bfaad645e2e33f; double * p = &x;}
  cint> p p
  (double*)0x100809480
  cint> p *p
  (double)(-8.31870270613823687e-286)

Unfortunately it is as slow as a dog. I'm not sure I understand what the pressing need for a C interpreter that is slow is. If so, there are two of them out there, including cint. The fact is, there is absolutely no technological reason why a C interpreter has to much slower than compiled C. Even the most stupid, naive implementation you can come up with is still as fast as Python (I know because I tried this last year). And with a proper Jit, it is absolutely possible to make it almost the same speed as compiled C.