I don't understand how this never got priority since 2015.
When providing Micropython API to an existing library, I have to duplicate all methods of base classes for each sub class instead of simply specify the base class. I get huge locals_dicts with duplicated data. Feels like the mechanism for extending Micropython is crippled.

I have to duplicate all methods of base classes

Not really, there are basic tricks which minimize boilerplate to the point you can't really speak of duplication anymore. For example:

def CopyAttributes(src, dst, names=None):
  if not names:
    names = (e for e in dir(src) if not e.startswith('__'))
  for entry in names:
    setattr(dst, entry, getattr(src, entry))

class Derived:
  def __init__(self):
    self.base = Base()
    CopyAttributes(self.base, self)

@stinos I'm not providing python code. I'm providing an extension to micropython by adding a module in C.
Do you suggest I post-process every object I add from C with python code? This isn't very scalable or convenient. May also have performance/memory penalty.

The same principle can also be applied in C if you're not using a const dict.

@stinos Please explain how to achieve this when adding a module. I tried setting .parent field of mp_obj_type_t but this does not work for native (C) objects.

I don't have an example ready but thinking of it, doing this in C is actually a bit different but less bloated as you don't need storage for each method. The idea is: you implement the type's attr function and when MicorPython asks you for an attribute you look it up in your class first, and if not found you look in your 'base' class instead and return it if found. Here's some untested pseudocode thrown together:

static void mytype_make_new(...) {
  //all the usual stuff, plus keep a 'base' class object
  self->myBasePtr = mybase_make_new(...);
}

static void myattr(mp_obj_t self_in, qstr attr, mp_obj_t* dest) {
  mytype *self = (mytype*) self_in;
  if(*dest == MP_OBJ_NULL) { //called to load attribute
      mp_obj_t myOwnAttribute = mp_map_lookup(self->base.type->locals_dict->map, new_qstr(attr), MP_MAP_LOOKUP);
      if(myOwnAttribute != MP_OBJ_NULL) {
        dest[0] = elem->value;
        dest[1] = self_in;
      }
      //we don't have that attribute, maybe base does? just forward call
    self->myBasePtr->base.type.attr(self->myBasePtr, attr, dest);
  } else {
    //do whatever needed to store attribute
  }
}

const mp_obj_type_t my_type = {
    //the usual stuff
    .make_new = mytype_make_new,
    .attr = myattr
};

@stinos Thanks for your example, I'll give it a try.
Don't you think this is something worth supporting as part of Micropython, instead of everyone implement the same functionality again and again?
Also I guess that in REPL, these functions will not be completed with TAB, will not appear in help etc...

Apologies if I'm missing the point here but it seems to me that providing generalised inheritance of built in types is considerably more complex than copying attributes. Consider properties like mutability, the ability to support item assignment, the ability to iterate over the instance, support for the buffer protocol and doubtless more. These may or may not be present in the base class.

@stinos @peterhinch It's worth noting that at least in Python this is already covered. (PEP253, an example)

Don't you think this is something worth supporting as part of Micropython, instead of everyone implement the same functionality again and again?

Yes of course, that's also why this issue exists.

Apologies if I'm missing the point here but it seems to me that providing generalised inheritance of built in types is considerably more complex

You're not missing any point, the example shown isn't generalised at all.

in Python this is already covered

Well I surely hope there are shortcuts not shown in those example then, instead of having to copy that 40-line type definition :P

See #4360 for a solution to this issue. It just does single inheritance but I think that is enough.