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/cmake/SuperBuild.cmake b/cmake/SuperBuild.cmake
index 8b895cc7..3bbda589 100644
--- a/cmake/SuperBuild.cmake
+++ b/cmake/SuperBuild.cmake
@@ -89,10 +89,12 @@ if(BUILD_OCC)
   set(OCC_DIR ${CMAKE_CURRENT_BINARY_DIR}/dependencies/occ)
 
   ExternalProject_Add(project_occ
-    URL https://github.com/Open-Cascade-SAS/OCCT/archive/refs/tags/V7_6_3.zip
-    URL_MD5 2426e373903faabbd4f96a01a934b66d
-    # URL https://github.com/Open-Cascade-SAS/OCCT/archive/refs/tags/V7_8_0.zip
-    # URL_MD5 f4432df8e42cb6178ea09a7448427f6c
+    # URL https://github.com/Open-Cascade-SAS/OCCT/archive/refs/tags/V7_6_3.zip
+    # URL_MD5 2426e373903faabbd4f96a01a934b66d
+    # URL https://github.com/Open-Cascade-SAS/OCCT/archive/refs/tags/V7_7_2.zip
+    # URL_MD5 533eb4f18af0f77ae321b158caeaee79
+    URL https://github.com/Open-Cascade-SAS/OCCT/archive/refs/tags/V7_8_1.zip
+    URL_MD5 bf62952a03696dab9e4272aa8efacb1a
     DOWNLOAD_DIR ${CMAKE_CURRENT_SOURCE_DIR}/external_dependencies
     ${SUBPROJECT_ARGS}
     CMAKE_ARGS
diff --git a/libsrc/core/CMakeLists.txt b/libsrc/core/CMakeLists.txt
index eca94b8c..6b60875d 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 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/generate_mpi_sources.py b/libsrc/core/generate_mpi_sources.py
index 1a491065..8c1e4db3 100644
--- a/libsrc/core/generate_mpi_sources.py
+++ b/libsrc/core/generate_mpi_sources.py
@@ -40,9 +40,11 @@ functions = [
         ("int", "MPI_Wait", "MPI_Request*", "MPI_Status*"),
         ("int", "MPI_Waitall", "int", "MPI_Request*:0", "MPI_Status*"),
         ("int", "MPI_Waitany", "int", "MPI_Request*:0", "int*", "MPI_Status*"),
+        ("int", "MPI_Comm_c2f", "MPI_Comm"),
         ]
 
 constants = [
+        ("MPI_Comm", "MPI_COMM_NULL"),
         ("MPI_Comm", "MPI_COMM_WORLD"),
         ("MPI_Datatype", "MPI_CHAR"),
         ("MPI_Datatype", "MPI_CXX_DOUBLE_COMPLEX"),
diff --git a/libsrc/core/mpi4py_pycapi.h b/libsrc/core/mpi4py_pycapi.h
new file mode 100644
index 00000000..87d0c69e
--- /dev/null
+++ b/libsrc/core/mpi4py_pycapi.h
@@ -0,0 +1,245 @@
+/* Author:  Lisandro Dalcin   */
+/* Contact: dalcinl@gmail.com */
+
+#ifndef MPI4PY_PYCAPI_H
+#define MPI4PY_PYCAPI_H
+
+#include <mpi.h>
+#include <Python.h>
+
+#define _mpi4py_declare_pycapi(Type, star) \
+static PyTypeObject *_mpi4py_PyMPI##Type = NULL; \
+static PyObject *(*_mpi4py_PyMPI##Type##_New)(MPI_##Type star) = NULL; \
+static MPI_##Type *(*_mpi4py_PyMPI##Type##_Get)(PyObject *) = NULL;
+
+#ifndef MPI4PY_LIMITED_API_SKIP_DATATYPE
+_mpi4py_declare_pycapi(Datatype,)
+#define PyMPIDatatype_Type (*_mpi4py_PyMPIDatatype)
+#define PyMPIDatatype_New _mpi4py_PyMPIDatatype_New
+#define PyMPIDatatype_Get _mpi4py_PyMPIDatatype_Get
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_STATUS
+_mpi4py_declare_pycapi(Status,*)
+#define PyMPIStatus_Type (*_mpi4py_PyMPIStatus)
+#define PyMPIStatus_New _mpi4py_PyMPIStatus_New
+#define PyMPIStatus_Get _mpi4py_PyMPIStatus_Get
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_REQUEST
+_mpi4py_declare_pycapi(Request,)
+#define PyMPIRequest_Type (*_mpi4py_PyMPIRequest)
+#define PyMPIRequest_New _mpi4py_PyMPIRequest_New
+#define PyMPIRequest_Get _mpi4py_PyMPIRequest_Get
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_MESSAGE
+_mpi4py_declare_pycapi(Message,)
+#define PyMPIMessage_Type (*_mpi4py_PyMPIMessage)
+#define PyMPIMessage_New _mpi4py_PyMPIMessage_New
+#define PyMPIMessage_Get _mpi4py_PyMPIMessage_Get
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_OP
+_mpi4py_declare_pycapi(Op,)
+#define PyMPIOp_Type (*_mpi4py_PyMPIOp)
+#define PyMPIOp_New _mpi4py_PyMPIOp_New
+#define PyMPIOp_Get _mpi4py_PyMPIOp_Get
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_GROUP
+_mpi4py_declare_pycapi(Group,)
+#define PyMPIGroup_Type (*_mpi4py_PyMPIGroup)
+#define PyMPIGroup_New _mpi4py_PyMPIGroup_New
+#define PyMPIGroup_Get _mpi4py_PyMPIGroup_Get
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_INFO
+_mpi4py_declare_pycapi(Info,)
+#define PyMPIInfo_Type (*_mpi4py_PyMPIInfo)
+#define PyMPIInfo_New _mpi4py_PyMPIInfo_New
+#define PyMPIInfo_Get _mpi4py_PyMPIInfo_Get
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_ERRHANDLER
+_mpi4py_declare_pycapi(Errhandler,)
+#define PyMPIErrhandler_Type (*_mpi4py_PyMPIErrhandler)
+#define PyMPIErrhandler_New _mpi4py_PyMPIErrhandler_New
+#define PyMPIErrhandler_Get _mpi4py_PyMPIErrhandler_Get
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_SESSION
+_mpi4py_declare_pycapi(Session,)
+#define PyMPISession_Type (*_mpi4py_PyMPISession)
+#define PyMPISession_New _mpi4py_PyMPISession_New
+#define PyMPISession_Get _mpi4py_PyMPISession_Get
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_COMM
+_mpi4py_declare_pycapi(Comm,)
+#define PyMPIComm_Type (*_mpi4py_PyMPIComm)
+#define PyMPIComm_New _mpi4py_PyMPIComm_New
+#define PyMPIComm_Get _mpi4py_PyMPIComm_Get
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_WIN
+_mpi4py_declare_pycapi(Win,)
+#define PyMPIWin_Type (*_mpi4py_PyMPIWin)
+#define PyMPIWin_New _mpi4py_PyMPIWin_New
+#define PyMPIWin_Get _mpi4py_PyMPIWin_Get
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_FILE
+_mpi4py_declare_pycapi(File,)
+#define PyMPIFile_Type (*_mpi4py_PyMPIFile)
+#define PyMPIFile_New _mpi4py_PyMPIFile_New
+#define PyMPIFile_Get _mpi4py_PyMPIFile_Get
+#endif
+
+#undef _mpi4py_define_pycapi
+
+static int _mpi4py_ImportType(PyObject *module,
+                              const char *type_name,
+                              PyTypeObject **type)
+{
+  PyObject *attr = NULL;
+  attr = PyObject_GetAttrString(module, type_name);
+  if (!attr)
+    goto fn_fail;
+  if (!PyType_Check(attr)) {
+    PyErr_Format(PyExc_TypeError,
+                 "%.200s.%.200s is not a type object",
+                 PyModule_GetName(module), type_name);
+    goto fn_fail;
+  }
+  *type = (PyTypeObject *)attr;
+  return 0;
+ fn_fail:
+  Py_DecRef(attr);
+  return -1;
+}
+
+static int _mpi4py_ImportFunc(PyObject *module,
+                              const char *func_name,
+                              const char *signature,
+                              void (**func)(void))
+{
+    PyObject *pyxcapi = NULL;
+    PyObject *capsule = NULL;
+    union { void *obj; void (*fcn)(void); } ptr;
+    pyxcapi = PyObject_GetAttrString(module, (char *)"__pyx_capi__");
+    if (!pyxcapi)
+      goto fn_fail;
+
+    capsule = PyDict_GetItemString(pyxcapi, func_name);
+    if (!capsule) {
+      PyErr_Format(PyExc_ImportError,
+                   "%.200s does not export expected C function %.200s",
+                   PyModule_GetName(module), func_name);
+      goto fn_fail;
+    }
+    if (!PyCapsule_CheckExact(capsule)) {
+      PyErr_Format(PyExc_TypeError,
+                   "%.200s.%.200s is not a capsule",
+                   PyModule_GetName(module), func_name);
+    }
+    if (!signature) {
+      signature = PyCapsule_GetName(capsule);
+    }
+    if (!PyCapsule_IsValid(capsule, signature)) {
+      PyErr_Format(PyExc_TypeError,
+                   "C function %.200s.%.200s has wrong signature "
+                   "(expected %.500s, got %.500s)",
+                   PyModule_GetName(module), func_name,
+                   signature, PyCapsule_GetName(capsule));
+      goto fn_fail;
+    }
+    ptr.obj = PyCapsule_GetPointer(capsule, signature);
+    if (!ptr.obj)
+      goto fn_fail;
+    *func = ptr.fcn;
+    Py_DecRef(pyxcapi);
+    return 0;
+ fn_fail:
+    Py_DecRef(pyxcapi);
+    return -1;
+}
+
+static int import_mpi4py_MPI(void)
+{
+  PyObject *module = PyImport_ImportModule("mpi4py.MPI");
+  if (!module)
+    goto fn_fail;
+
+#define _mpi4py_import_pycapi(Type) do {                                     \
+    if (_mpi4py_ImportType(module, #Type, &_mpi4py_PyMPI##Type) < 0)         \
+      goto fn_fail;                                                          \
+    if (_mpi4py_ImportFunc(module, "PyMPI" #Type "_New", NULL,               \
+                           (void (**)(void))&_mpi4py_PyMPI##Type##_New) < 0) \
+      goto fn_fail;                                                          \
+    if (_mpi4py_ImportFunc(module, "PyMPI" #Type "_Get", NULL,               \
+                           (void (**)(void))&_mpi4py_PyMPI##Type##_Get) < 0) \
+      goto fn_fail;                                                          \
+  } while (0)
+
+#ifndef MPI4PY_LIMITED_API_SKIP_DATATYPE
+  _mpi4py_import_pycapi(Datatype);
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_STATUS
+  _mpi4py_import_pycapi(Status);
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_REQUEST
+  _mpi4py_import_pycapi(Request);
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_MESSAGE
+  _mpi4py_import_pycapi(Message);
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_OP
+  _mpi4py_import_pycapi(Op);
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_GROUP
+  _mpi4py_import_pycapi(Group);
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_INFO
+  _mpi4py_import_pycapi(Info);
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_ERRHANDLER
+  _mpi4py_import_pycapi(Errhandler);
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_SESSION
+  _mpi4py_import_pycapi(Session);
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_COMM
+  _mpi4py_import_pycapi(Comm);
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_WIN
+  _mpi4py_import_pycapi(Win);
+#endif
+
+#ifndef MPI4PY_LIMITED_API_SKIP_FILE
+  _mpi4py_import_pycapi(File);
+#endif
+
+#undef _mpi4py_import_pycapi
+
+  Py_DecRef(module);
+  return 0;
+ fn_fail:
+  Py_DecRef(module);
+  return -1;
+}
+
+#define __PYX_HAVE_API__mpi4py__MPI
+#define import_mpi4py__MPI import_mpi4py_MPI
+
+#endif /* MPI4PY_PYCAPI_H */
diff --git a/libsrc/core/ng_mpi.cpp b/libsrc/core/ng_mpi.cpp
index 9dce98c8..bac75e47 100644
--- a/libsrc/core/ng_mpi.cpp
+++ b/libsrc/core/ng_mpi.cpp
@@ -1,22 +1,23 @@
 #define OMPI_SKIP_MPICXX
+#include <mpi.h>
 
 #include "ng_mpi.hpp"
 
-#include <mpi.h>
-
 #include <type_traits>
 
+#include "array.hpp"
 #include "ngcore_api.hpp"
 #include "pybind11/pytypes.h"
 
-#if defined(NG_PYTHON) && defined(NG_MPI4PY)
-#include <mpi4py.h>
-
+#ifdef NG_PYTHON
 #include "python_ngcore.hpp"
-
-namespace py = pybind11;
 #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
+
 #ifdef MSMPI_VER
 int MPI_Comm_create_group(MPI_Comm arg0, MPI_Group arg1, int arg2,
                           MPI_Comm* arg3) {
@@ -156,10 +157,10 @@ 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();
@@ -172,12 +173,12 @@ 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..36151c09 100644
--- a/libsrc/core/ng_mpi.hpp
+++ b/libsrc/core/ng_mpi.hpp
@@ -9,17 +9,9 @@
 
 #include "ngcore_api.hpp"
 
-#if defined(NG_PYTHON) && defined(NG_MPI4PY)
-#include <pybind11/pybind11.h>
-
-namespace py = pybind11;
-#endif
-
 #ifndef NG_MPI_WRAPPER
+#define OMPI_SKIP_MPICXX
 #include <mpi.h>
-#if defined(NG_PYTHON) && defined(NG_MPI4PY)
-#include <mpi4py.h>
-#endif
 #endif  // NG_MPI_WRAPPER
 
 namespace ngcore {
@@ -83,23 +75,18 @@ 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_generated_declarations.hpp b/libsrc/core/ng_mpi_generated_declarations.hpp
index d6a9e33f..870ed34f 100644
--- a/libsrc/core/ng_mpi_generated_declarations.hpp
+++ b/libsrc/core/ng_mpi_generated_declarations.hpp
@@ -40,6 +40,8 @@ NGCORE_API extern int (*NG_MPI_Type_size)(NG_MPI_Datatype, int*);
 NGCORE_API extern int (*NG_MPI_Wait)(NG_MPI_Request*, NG_MPI_Status*);
 NGCORE_API extern int (*NG_MPI_Waitall)(int, NG_MPI_Request*, NG_MPI_Status*);
 NGCORE_API extern int (*NG_MPI_Waitany)(int, NG_MPI_Request*, int*, NG_MPI_Status*);
+NGCORE_API extern int (*NG_MPI_Comm_c2f)(NG_MPI_Comm);
+NGCORE_API extern NG_MPI_Comm NG_MPI_COMM_NULL;
 NGCORE_API extern NG_MPI_Comm NG_MPI_COMM_WORLD;
 NGCORE_API extern NG_MPI_Datatype NG_MPI_CHAR;
 NGCORE_API extern NG_MPI_Datatype NG_MPI_CXX_DOUBLE_COMPLEX;
@@ -107,6 +109,8 @@ static const auto NG_MPI_Type_size = MPI_Type_size;
 static const auto NG_MPI_Wait = MPI_Wait;
 static const auto NG_MPI_Waitall = MPI_Waitall;
 static const auto NG_MPI_Waitany = MPI_Waitany;
+static const auto NG_MPI_Comm_c2f = MPI_Comm_c2f;
+static const decltype(MPI_COMM_NULL) NG_MPI_COMM_NULL = MPI_COMM_NULL;
 static const decltype(MPI_COMM_WORLD) NG_MPI_COMM_WORLD = MPI_COMM_WORLD;
 static const decltype(MPI_CHAR) NG_MPI_CHAR = MPI_CHAR;
 static const decltype(MPI_CXX_DOUBLE_COMPLEX) NG_MPI_CXX_DOUBLE_COMPLEX = MPI_CXX_DOUBLE_COMPLEX;
diff --git a/libsrc/core/ng_mpi_generated_dummy_init.hpp b/libsrc/core/ng_mpi_generated_dummy_init.hpp
index 6050bdee..9ea6fced 100644
--- a/libsrc/core/ng_mpi_generated_dummy_init.hpp
+++ b/libsrc/core/ng_mpi_generated_dummy_init.hpp
@@ -39,6 +39,8 @@ decltype(NG_MPI_Type_size) NG_MPI_Type_size = [](NG_MPI_Datatype, int*)->int { t
 decltype(NG_MPI_Wait) NG_MPI_Wait = [](NG_MPI_Request*, NG_MPI_Status*)->int { throw no_mpi(); };
 decltype(NG_MPI_Waitall) NG_MPI_Waitall = [](int, NG_MPI_Request*, NG_MPI_Status*)->int { throw no_mpi(); };
 decltype(NG_MPI_Waitany) NG_MPI_Waitany = [](int, NG_MPI_Request*, int*, NG_MPI_Status*)->int { throw no_mpi(); };
+decltype(NG_MPI_Comm_c2f) NG_MPI_Comm_c2f = [](NG_MPI_Comm)->int { throw no_mpi(); };
+NG_MPI_Comm NG_MPI_COMM_NULL = 0;
 NG_MPI_Comm NG_MPI_COMM_WORLD = 0;
 NG_MPI_Datatype NG_MPI_CHAR = 0;
 NG_MPI_Datatype NG_MPI_CXX_DOUBLE_COMPLEX = 0;
diff --git a/libsrc/core/ng_mpi_generated_init.hpp b/libsrc/core/ng_mpi_generated_init.hpp
index b44bc2a7..2529f409 100644
--- a/libsrc/core/ng_mpi_generated_init.hpp
+++ b/libsrc/core/ng_mpi_generated_init.hpp
@@ -39,6 +39,8 @@ NG_MPI_Type_size = [](NG_MPI_Datatype arg0, int* arg1)->int { return MPI_Type_si
 NG_MPI_Wait = [](NG_MPI_Request* arg0, NG_MPI_Status* arg1)->int { return MPI_Wait( ng2mpi(arg0),  ng2mpi(arg1)); };
 NG_MPI_Waitall = [](int arg0, NG_MPI_Request* arg1, NG_MPI_Status* arg2)->int { return MPI_Waitall( arg0,  ng2mpi(arg1, arg0),  ng2mpi(arg2)); };
 NG_MPI_Waitany = [](int arg0, NG_MPI_Request* arg1, int* arg2, NG_MPI_Status* arg3)->int { return MPI_Waitany( arg0,  ng2mpi(arg1, arg0),  arg2,  ng2mpi(arg3)); };
+NG_MPI_Comm_c2f = [](NG_MPI_Comm arg0)->int { return MPI_Comm_c2f( ng2mpi(arg0)); };
+NG_MPI_COMM_NULL = mpi2ng(MPI_COMM_NULL);
 NG_MPI_COMM_WORLD = mpi2ng(MPI_COMM_WORLD);
 NG_MPI_CHAR = mpi2ng(MPI_CHAR);
 NG_MPI_CXX_DOUBLE_COMPLEX = mpi2ng(MPI_CXX_DOUBLE_COMPLEX);
diff --git a/libsrc/core/ng_mpi_native.hpp b/libsrc/core/ng_mpi_native.hpp
deleted file mode 100644
index 6c8f40ce..00000000
--- a/libsrc/core/ng_mpi_native.hpp
+++ /dev/null
@@ -1,21 +0,0 @@
-#ifndef NG_MPI_NATIVE_HPP
-#define NG_MPI_NATIVE_HPP
-
-#include <mpi.h>
-
-#include "mpi_wrapper.hpp"
-#include "ng_mpi.hpp"
-
-namespace ngcore {
-
-MPI_Comm NG_MPI_Native(NG_MPI_Comm comm) {
-  return reinterpret_cast<MPI_Comm>(comm.value);
-}
-
-MPI_Comm NG_MPI_Native(NgMPI_Comm comm) {
-  return reinterpret_cast<MPI_Comm>(static_cast<NG_MPI_Comm>(comm).value);
-}
-
-}  // namespace ngcore
-
-#endif  // NG_MPI_NATIVE_HPP
diff --git a/libsrc/core/ng_mpi_wrapper.cpp b/libsrc/core/ng_mpi_wrapper.cpp
index ac9d3894..835b0c31 100644
--- a/libsrc/core/ng_mpi_wrapper.cpp
+++ b/libsrc/core/ng_mpi_wrapper.cpp
@@ -13,6 +13,13 @@ using std::cerr;
 using std::cout;
 using std::endl;
 
+#ifndef NG_MPI_WRAPPER
+#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 +35,7 @@ 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 +133,7 @@ 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 +157,17 @@ 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 +182,19 @@ 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
 }
diff --git a/libsrc/meshing/CMakeLists.txt b/libsrc/meshing/CMakeLists.txt
index 8ca9a0cc..697e8ad1 100644
--- a/libsrc/meshing/CMakeLists.txt
+++ b/libsrc/meshing/CMakeLists.txt
@@ -15,7 +15,7 @@ target_sources(nglib PRIVATE
         boundarylayer2d.cpp
 )
 
-target_link_libraries( nglib PRIVATE netgen_metis "$<BUILD_INTERFACE:netgen_python>" )
+target_link_libraries( nglib PRIVATE $<BUILD_INTERFACE:netgen_metis> $<BUILD_INTERFACE:netgen_python> )
 
 install(FILES
   adfront2.hpp adfront3.hpp basegeom.hpp bcfunctions.hpp bisect.hpp
diff --git a/libsrc/meshing/python_mesh.cpp b/libsrc/meshing/python_mesh.cpp
index f93ed62f..3a652e02 100644
--- a/libsrc/meshing/python_mesh.cpp
+++ b/libsrc/meshing/python_mesh.cpp
@@ -87,6 +87,8 @@ DLL_HEADER void ExportNetgenMeshing(py::module &m)
   py::class_<NGDummyArgument>(m, "NGDummyArgument")
     .def("__bool__", []( NGDummyArgument &self ) { return false; } )
     ;
+
+  py::class_<LocalH, shared_ptr<LocalH>>(m, "LocalH");
   
   py::class_<Point<2>> (m, "Point2d")
     .def(py::init<double,double>())
@@ -1249,7 +1251,11 @@ DLL_HEADER void ExportNetgenMeshing(py::module &m)
             else mp.optsteps3d = 5;
             OptimizeVolume (mp, self);
           }, py::arg("mp"), py::call_guard<py::gil_scoped_release>())
-
+    .def("SetLocalH",[](Mesh& self, shared_ptr<LocalH> localh, int layer)
+         {
+           self.SetLocalH(localh, layer);
+         }, py::arg("localh"), py::arg("layer")=1)
+    .def("GetLocalH", &Mesh::GetLocalH)
     .def ("OptimizeMesh2d", [](Mesh & self, MeshingParameters* pars, int faceindex)
           {
             self.CalcLocalH(0.5);
diff --git a/setup.py b/setup.py
index 90b06d0f..80371bab 100644
--- a/setup.py
+++ b/setup.py
@@ -120,6 +120,7 @@ elif 'linux' in sys.platform:
 if have_mpi:
     cmake_args += [
         '-DUSE_MPI=ON',
+        '-DUSE_MPI_WRAPPER=ON',
     ]
 
 cmake_args += [