Follow-up with numbers, as promised. Most of the cost isn't the routes. The first registered route costs 5-9 s and each route after that about 0.1 s, whatever the handler signature. The fixed part is the session's Beast/Asio read and write path, which is the same in every program. The separately compiled mode is now in master: define BOOST_TAAR_SEPARATE_COMPILATION and include <boost/taar/src.hpp> in one TU, the same pattern as Asio and Beast. Same examples, debug, GCC 16 / Clang 23: hello_server 13.2 -> 4.6 s / 8.1 -> 3.7 s rest_crud_server 18.4 -> 9.2 s / 10.7 -> 6.2 s mcp_server 20.7 -> 9.6 s / 12.7 -> 6.9 s What's left is mostly parsing the Asio/Beast headers plus the per-handler binding, which has to stay templates. The trade-off is binary size: release builds come out 0.3-1.1 MB larger (unused instantiations), much of which --gc-sections recovers. All tests pass in both modes, and CI covers both on GCC, Clang and MSVC. Header-only stays the default. The numbers and usage are in the docs: https://rjahanbakhshi.github.io/boost.taar/boost_taar/quickstart.html#boost_... Artyom, on static polymorphism: the business logic doesn't have to be templated. Handlers are plain functions and can live in your own .cpp files without including taar; only the file that registers routes sees the templates. With separate compilation, the HTTP and HTTPS sessions are each compiled once, in one file. Reza