For the SOGo/OpenChange project, we are now using your converter for interpreting RTF content. In order to do so, I have modified the code of unrtf in order to let it built as a library. Mainly, this implied converting all global variables into members of structures that are passed from function to function. The structures in questions have been layed out as this: one for the conversion parameters inherited from the command line, another for all internal processing as well as the pre-existing one for the output personality.
I have not made changes to the existing code per se except for fixing a few leak here and there, however I have added an interface header (unrtf.h) with the public conversion functions, that main.c now invokes. Also, I have modified the code in parser.c to use a string as input rather than a FILE * and improved the "safe_printf" to make use of a configurable output method while keeping its safety purpose. Consequently, the public function for file handling first load the entire file in memory before passing its content to the string-based function, and the output can now be directed to a file, to screen or to another structure provided by the client of the library.

Besides all of that, I have also reindented most of the code in a way that's more compliant to the GNU coding standards, as it's my default editing setup for my projects. If it bothers you, I could probably produce another patch that ignores changes to whitespaces. I also remove the path.c module from the build as it was no longer required.

For the SOGo/OpenChange backend, I am using a special makefile that integrates with the GNUstep-based SOGo build. Therefore I made no modifications to the current build scripts of unrtf, besides removing path.c from the corresponding Makefile.am. This means that even though it's ready to be built as a library, it is not when built from the current scripts.

If you decide to go further with this, I think my patch is a good starting point. I can see other things that could be done beforehand: fix the code syntax to be more compliant with the GNU coding standards (with regards to the layout of braces for example), replace the loading of the input file with a mmap-based mechanism, in order to reduce memory usage for big files.
Also, the "ConversionContext" structure could be split in multiple substructures in order to organize its members.