diff --git a/CMakeLists.txt b/CMakeLists.txt
index 71ac2d86..31faba7e 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -16,8 +16,7 @@ option( USE_GUI     "build with GUI" ON )
 option( USE_PYTHON  "build with python interface" ON )
 cmake_dependent_option( PREFER_SYSTEM_PYBIND11  "Use system wide PyBind11" OFF "USE_PYTHON" OFF)
 option( USE_MPI     "enable mpi parallelization" OFF )
-option( USE_MPI4PY  "enable mpi4py interface" ON )
-option( USE_MPI_WRAPPER "enable mpi wrapper (run-time dispatch of MPI library calls)" ON )
+option( USE_MPI_WRAPPER "enable mpi wrapper (run-time dispatch of MPI library calls)" OFF )
 option( USE_OCC     "build with OpenCascade geometry kernel interface" ON)
 option( USE_STLGEOM     "build with STL geometry support" ON)
 option( USE_CSG     "build with CSG kernel" ON)
@@ -323,6 +322,7 @@ if (USE_PYTHON)
       add_subdirectory(external_dependencies/pybind11)
     endif()
 
+    target_compile_definitions(netgen_python INTERFACE NG_PYTHON NETGEN_PYTHON)
     target_include_directories(netgen_python INTERFACE ${pybind11_INCLUDE_DIR} ${Python3_INCLUDE_DIRS})
     target_include_directories(nglib PRIVATE ${pybind11_INCLUDE_DIR} ${Python3_INCLUDE_DIRS})
     if(Python3_LIBRARIES AND (WIN32 OR NOT BUILD_FOR_CONDA))
@@ -345,14 +345,6 @@ if (USE_MPI)
     target_include_directories(netgen_metis INTERFACE ${METIS_INCLUDE_DIR})
     target_link_libraries(netgen_metis INTERFACE ${METIS_LIBRARY} )
     target_compile_definitions(netgen_metis INTERFACE METIS )
-
-    if(USE_MPI4PY AND USE_PYTHON)
-      execute_process(COMMAND ${Python3_EXECUTABLE} -c "import mpi4py;print(mpi4py.get_include())" OUTPUT_VARIABLE mpi4py_path OUTPUT_STRIP_TRAILING_WHITESPACE)
-      find_path(MPI4PY_INCLUDE_DIR mpi4py.h HINTS ${mpi4py_path}/mpi4py NO_DEFAULT_PATH REQUIRED)
-      target_include_directories(netgen_python INTERFACE ${MPI4PY_INCLUDE_DIR})
-      target_compile_definitions(netgen_python INTERFACE NG_MPI4PY )
-      message(STATUS "Found mpi4py: ${MPI4PY_INCLUDE_DIR}")
-    endif(USE_MPI4PY AND USE_PYTHON)
 endif (USE_MPI)
 
 #######################################################################
diff --git a/libsrc/core/CMakeLists.txt b/libsrc/core/CMakeLists.txt
index eca94b8c..69245779 100644
--- a/libsrc/core/CMakeLists.txt
+++ b/libsrc/core/CMakeLists.txt
@@ -93,7 +93,7 @@ install(FILES ngcore.hpp archive.hpp type_traits.hpp version.hpp ngcore_api.hpp
   xbool.hpp signal.hpp bitarray.hpp table.hpp hashtable.hpp ranges.hpp ngstream.hpp
   simd.hpp simd_avx.hpp simd_avx512.hpp simd_generic.hpp simd_sse.hpp simd_arm64.hpp
   register_archive.hpp autodiff.hpp autodiffdiff.hpp
-  ng_mpi.hpp ng_mpi_generated_declarations.hpp ng_mpi_native.hpp
+  ng_mpi.hpp ng_mpi_generated_declarations.hpp ng_mpi_native.hpp mpi4py_pycapi.h
   DESTINATION ${NG_INSTALL_DIR_INCLUDE}/core COMPONENT netgen_devel)
 
 if(ENABLE_CPP_CORE_GUIDELINES_CHECK)
@@ -166,6 +166,7 @@ if(USE_MPI)
     endif()
   else()
     target_link_libraries(ngcore PUBLIC ${MPI_C_LIBRARIES})
+    target_include_directories(ngcore PUBLIC ${MPI_C_INCLUDE_PATH})
   endif(USE_MPI_WRAPPER)
 
 endif(USE_MPI)
diff --git a/libsrc/core/ng_mpi.cpp b/libsrc/core/ng_mpi.cpp
index 30075ed0..0cb2847f 100644
--- a/libsrc/core/ng_mpi.cpp
+++ b/libsrc/core/ng_mpi.cpp
@@ -6,20 +6,18 @@
 
 #include <type_traits>
 
+#include "array.hpp"
 #include "ngcore_api.hpp"
 #include "pybind11/pytypes.h"
 
-#if defined(NG_PYTHON) && defined(NG_MPI4PY)
+#ifdef NG_PYTHON
+#include "python_ngcore.hpp"
+#endif
+
 #define MPI4PY_LIMITED_API 1
 #define MPI4PY_LIMITED_API_SKIP_MESSAGE 1
 #define MPI4PY_LIMITED_API_SKIP_SESSION 1
-#include "mpi4py_pycapi.h" // mpi4py < 4.0.0
-#include <mpi4py.h>
-
-#include "python_ngcore.hpp"
-
-namespace py = pybind11;
-#endif
+#include "mpi4py_pycapi.h"  // mpi4py < 4.0.0
 
 #ifdef MSMPI_VER
 int MPI_Comm_create_group(MPI_Comm arg0, MPI_Group arg1, int arg2,
@@ -160,10 +158,10 @@ NGCORE_API_EXPORT void ng_init_mpi();
 static bool imported_mpi4py = false;
 
 void ng_init_mpi() {
-#if defined(NG_PYTHON) && defined(NG_MPI4PY)
+#ifdef NG_PYTHON
   NG_MPI_CommFromMPI4Py = [](py::handle src, NG_MPI_Comm& dst) -> bool {
     if (!imported_mpi4py) {
-      import_mpi4py();
+      import_mpi4py__MPI();
       imported_mpi4py = true;
     }
     PyObject* py_src = src.ptr();
@@ -176,12 +174,12 @@ void ng_init_mpi() {
   };
   NG_MPI_CommToMPI4Py = [](NG_MPI_Comm src) -> py::handle {
     if (!imported_mpi4py) {
-      import_mpi4py();
+      import_mpi4py__MPI();
       imported_mpi4py = true;
     }
     return py::handle(PyMPIComm_New(ng2mpi(src)));
   };
-#endif
+#endif  // NG_PYTHON
 
 #include "ng_mpi_generated_init.hpp"
 }
diff --git a/libsrc/core/ng_mpi.hpp b/libsrc/core/ng_mpi.hpp
index f29bb664..9843a516 100644
--- a/libsrc/core/ng_mpi.hpp
+++ b/libsrc/core/ng_mpi.hpp
@@ -9,17 +9,8 @@
 
 #include "ngcore_api.hpp"
 
-#if defined(NG_PYTHON) && defined(NG_MPI4PY)
-#include <pybind11/pybind11.h>
-
-namespace py = pybind11;
-#endif
-
 #ifndef NG_MPI_WRAPPER
 #include <mpi.h>
-#if defined(NG_PYTHON) && defined(NG_MPI4PY)
-#include <mpi4py.h>
-#endif
 #endif  // NG_MPI_WRAPPER
 
 namespace ngcore {
@@ -83,23 +74,18 @@ struct NG_MPI_Aint {
   NG_MPI_Aint() = default;
 };
 
-#else
+#else   // NG_MPI_WRAPPER
+using NG_MPI_Comm = MPI_Comm;
 using NG_MPI_Status = MPI_Status;
-using NG_MPI_Comm  = MPI_Comm;
 using NG_MPI_Datatype = MPI_Datatype;
 using NG_MPI_Request = MPI_Request;
 using NG_MPI_Op = MPI_Op;
 using NG_MPI_Group = MPI_Group;
 using NG_MPI_Aint = MPI_Aint;
-#endif
+#endif  // NG_MPI_WRAPPER
 
 #include "ng_mpi_generated_declarations.hpp"
 
-#if defined(NG_PYTHON) && defined(NG_MPI4PY)
-NGCORE_API extern bool (*NG_MPI_CommFromMPI4Py)(py::handle, NG_MPI_Comm &);
-NGCORE_API extern py::handle (*NG_MPI_CommToMPI4Py)(NG_MPI_Comm);
-#endif
-
 }  // namespace ngcore
 
 #endif  // PARALLEL
diff --git a/libsrc/core/ng_mpi_wrapper.cpp b/libsrc/core/ng_mpi_wrapper.cpp
index ac9d3894..eedaf1b4 100644
--- a/libsrc/core/ng_mpi_wrapper.cpp
+++ b/libsrc/core/ng_mpi_wrapper.cpp
@@ -13,6 +13,14 @@ using std::cerr;
 using std::cout;
 using std::endl;
 
+#ifndef NG_MPI_WRAPPER
+#include <mpi.h>
+#define MPI4PY_LIMITED_API 1
+#define MPI4PY_LIMITED_API_SKIP_MESSAGE 1
+#define MPI4PY_LIMITED_API_SKIP_SESSION 1
+#include "mpi4py_pycapi.h"  // mpi4py < 4.0.0
+#endif // NG_MPI_WRAPPER
+
 namespace ngcore {
 
 #ifdef NG_MPI_WRAPPER
@@ -28,9 +36,7 @@ struct MPIFinalizer {
   }
 } mpi_finalizer;
 
-bool MPI_Loaded() {
-  return ng_mpi_lib != nullptr;
-}
+bool MPI_Loaded() { return ng_mpi_lib != nullptr; }
 
 void InitMPI(std::optional<std::filesystem::path> mpi_lib_path) {
   if (ng_mpi_lib) return;
@@ -128,7 +134,7 @@ static std::runtime_error no_mpi() {
   return std::runtime_error("MPI not enabled");
 }
 
-#if defined(NG_PYTHON) && defined(NG_MPI4PY)
+#ifdef NG_PYTHON
 decltype(NG_MPI_CommFromMPI4Py) NG_MPI_CommFromMPI4Py =
     [](py::handle py_obj, NG_MPI_Comm &ng_comm) -> bool {
   // If this gets called, it means that we want to convert an mpi4py
@@ -152,17 +158,17 @@ decltype(NG_MPI_CommFromMPI4Py) NG_MPI_CommFromMPI4Py =
 };
 decltype(NG_MPI_CommToMPI4Py) NG_MPI_CommToMPI4Py =
     [](NG_MPI_Comm) -> py::handle { throw no_mpi(); };
-#endif
+#endif  // NG_PYTHON
 
 #include "ng_mpi_generated_dummy_init.hpp"
 #else  // NG_MPI_WRAPPER
 
 static bool imported_mpi4py = false;
-#if defined(NG_PYTHON) && defined(NG_MPI4PY)
+#ifdef NG_PYTHON
 decltype(NG_MPI_CommFromMPI4Py) NG_MPI_CommFromMPI4Py =
     [](py::handle src, NG_MPI_Comm &dst) -> bool {
   if (!imported_mpi4py) {
-    import_mpi4py();
+    import_mpi4py__MPI();
     imported_mpi4py = true;
   }
   PyObject *py_src = src.ptr();
@@ -177,19 +183,19 @@ decltype(NG_MPI_CommFromMPI4Py) NG_MPI_CommFromMPI4Py =
 decltype(NG_MPI_CommToMPI4Py) NG_MPI_CommToMPI4Py =
     [](NG_MPI_Comm src) -> py::handle {
   if (!imported_mpi4py) {
-    import_mpi4py();
+    import_mpi4py__MPI();
     imported_mpi4py = true;
   }
   return py::handle(PyMPIComm_New(src));
 };
 
-#endif
+#endif  // NG_PYTHON
 
+bool MPI_Loaded() { return true; }
 void InitMPI(std::optional<std::filesystem::path>) {}
 
 #endif  // NG_MPI_WRAPPER
 
-
 }  // namespace ngcore
 
 #endif  // PARALLEL
diff --git a/libsrc/core/python_ngcore.hpp b/libsrc/core/python_ngcore.hpp
index cd3d14fb..c8ad53ac 100644
--- a/libsrc/core/python_ngcore.hpp
+++ b/libsrc/core/python_ngcore.hpp
@@ -13,13 +13,17 @@
 #include "archive.hpp"
 #include "flags.hpp"
 #include "ngcore_api.hpp"
-#include "profiler.hpp"
 #include "ng_mpi.hpp"
 
 namespace py = pybind11;
 
 namespace ngcore
 {
+#ifdef PARALLEL
+  NGCORE_API extern bool (*NG_MPI_CommFromMPI4Py)(py::handle, NG_MPI_Comm &);
+  NGCORE_API extern py::handle (*NG_MPI_CommToMPI4Py)(NG_MPI_Comm);
+#endif // PARALLEL
+
   namespace detail
   {
     template<typename T>
@@ -34,15 +38,15 @@ namespace ngcore
     };
   } // namespace detail
 
+#ifdef PARALLEL
   struct mpi4py_comm {
     mpi4py_comm() = default;
-#ifdef PARALLEL
     mpi4py_comm(NG_MPI_Comm value) : value(value) {}
     operator NG_MPI_Comm () { return value; }
 
     NG_MPI_Comm value;
-#endif  // PARALLEL
   };
+#endif  // PARALLEL
 } // namespace ngcore
 
 
@@ -51,7 +55,7 @@ namespace ngcore
 namespace pybind11 {
 namespace detail {
 
-#ifdef NG_MPI4PY
+#ifdef PARALLEL
 template <> struct type_caster<ngcore::mpi4py_comm> {
   public:
   PYBIND11_TYPE_CASTER(ngcore::mpi4py_comm, _("mpi4py_comm"));
@@ -70,7 +74,7 @@ template <> struct type_caster<ngcore::mpi4py_comm> {
       return ngcore::NG_MPI_CommToMPI4Py(src.value);
     }
 };
-#endif // NG_MPI4PY
+#endif //  PARALLEL
 
 template <typename Type, typename Value> struct ngcore_list_caster {
     using value_conv = make_caster<Value>;
diff --git a/libsrc/core/python_ngcore_export.cpp b/libsrc/core/python_ngcore_export.cpp
index 8c9e1a0e..fdcc4bb2 100644
--- a/libsrc/core/python_ngcore_export.cpp
+++ b/libsrc/core/python_ngcore_export.cpp
@@ -371,5 +371,7 @@ threads : int
   ;
 
     
+#ifdef PARALLEL
   py::implicitly_convertible<mpi4py_comm, NgMPI_Comm>();
+#endif // PARALLEL
 }