diff --git a/.gitignore b/.gitignore
index 2026309..8e1b87c 100644
--- a/.gitignore
+++ b/.gitignore
@@ -63,4 +63,5 @@ test_regsurf
 *.gcov
 gmon.out
 npytest
+*.npy
 # CVS default ignores end
diff --git a/t.cc b/t.cc
index 81212c0..322883d 100644
--- a/t.cc
+++ b/t.cc
@@ -159,7 +159,7 @@ cout <<y;
 
 
 
-int main()
+int main(int argc, char **argv)
 {
 sigtraceback(SIGSEGV,1);
 sigtraceback(SIGABRT,1);
@@ -4746,7 +4746,7 @@ cout <<aa;
 cout <<"Error = "<<(ax-aa).norm()<<endl;
 }
 
-if(1)
+if(0)
 {
 int n=3;
 NRVec<INDEXGROUP> s(4);
@@ -4766,5 +4766,36 @@ Tensor<double> t1 = t.subtensor1(2);
 cout <<t1;
 }
 
+if(0)
+{
+int n=2;
+NRVec<INDEXGROUP> s(4);
+for(int i=0; i<4; ++i) 
+	{
+	s[i].number=1;
+	s[i].symmetry=0;
+	s[i].offset=0;
+	s[i].range=n+i;
+	}
+Tensor<double> t(s);
+t.randomize(1.);
+t.numpy_write("test.npy","<f8");
+
+Tensor<double> tt;
+tt.numpy_read("test.npy");
+cout<<t;
+cout<<tt;
+cout <<"Error = "<<(t-tt).norm()<<endl;
+}
+
+if(1)
+{
+Tensor<double> tt;
+tt.numpy_read(argv[1]);
+//cout<<tt;
+cout<<"part\n"<<tt.subtensor1(0);
+cout<<"part\n"<<tt.subtensor1(1);
+cout<<"part\n"<<tt.subtensor1(2);
+}
 
 }//main
diff --git a/tensor.cc b/tensor.cc
index 66a7b4c..e8e5dc2 100644
--- a/tensor.cc
+++ b/tensor.cc
@@ -2514,6 +2514,110 @@ return r;
 }
 
 
+//NOTE: we do not check numpy type matches T
+template<typename T>
+void Tensor<T>::numpy_write(const char *name, const char *descr) const
+{
+if(!is_flat()) laerror("numpy_write only for flat tensors");
+int fd=open(name,O_CREAT|O_LARGEFILE|O_RDWR,0777);
+if(fd<0) laerror("cannot open in numpy_write");
+char magic[]="\x93NUMPY\x01\x00";
+if(8!=write(fd,magic,8)) laerror("cannot write 1 in numpy_write");
+
+//construct header
+char header[2048];
+int16_t header_len=0;
+sprintf(header,"{'descr': '%s', 'fortran_order': False, 'shape': (",descr);
+for(int i=myrank-1; i>=0; --i)
+	{
+	header_len=strlen(header);
+	sprintf(header+header_len,"%d",shape[i].range);
+	if(i>0)
+		{
+		header_len=strlen(header);
+		sprintf(header+header_len,", ");
+		}
+	}
+
+header_len=strlen(header);
+sprintf(header+header_len,"), }");
+
+//pad header by spaces to 8 byte boundary and terminate by \n
+header_len=strlen(header);
+int x= 8 - (header_len+2+8)%8;
+for(int i=0; i<x; ++i) header[header_len+i]=' ';
+header_len+=x;
+header[header_len-1]='\n';
+header[header_len]=0;
+
+if(2!=write(fd,&header_len,2)) laerror("cannot write 2 in numpy_write");
+if(header_len!=write(fd,header,header_len)) laerror("cannot write 3 in numpy_write");
+
+if(sizeof(T)*data.size()!= write(fd,&data[0],sizeof(T)*data.size())) laerror("cannot write 4 in numpy_write");
+
+if(close(fd)) laerror("cannot close in numpy_write");
+}
+
+
+
+//NOTE: we do not check numpy type matches T
+template<typename T>
+void Tensor<T>::numpy_read(const char *name)
+{
+int fd=open(name,O_LARGEFILE|O_RDONLY);
+if(fd<0) laerror("cannot open in numpy_read");
+char magic[]="\x93NUMPY\x01\x00";
+char readmagic[8];
+if(8!=read(fd,readmagic,8)) laerror("cannot read 1 in numpy_read");
+if(memcmp(magic,readmagic,8)) laerror("bad magic or version != 1.0 in numpy_read");
+
+int16_t header_len;
+if(2!=read(fd,&header_len,2)) laerror("cannot read 2 in numpy_read");
+
+char header[header_len];
+if(header_len!=read(fd,header,header_len)) laerror("cannot read 3 in numpy_read");
+
+char *p=strstr(header,"fortran_order");
+if(!p) laerror("cannot find fortran_order in numpy_read");
+bool fortranorder=false;
+p=strstr(p,": ");
+if(strncmp(p+2,"False",5)) fortranorder=true;
+//std::cout <<"Fortran order "<<fortranorder<<std::endl;
+
+std::list<int> dimslist;
+p=strstr(p,"'shape': (");
+if(!p) laerror("cannot find shape in numpy_read");
+p+=strlen("'shape': (");
+int d;
+while(1==sscanf(p,"%d",&d))
+	{
+	if(fortranorder) dimslist.push_back(d); else  dimslist.push_front(d);
+	while(isdigit(*p)||isspace(*p)) p++;
+	if(*p==')') break;
+	++p; //skip ,
+	}
+
+NRVec<int> dims(dimslist);
+if(dims.size()==0) laerror("zero-dimensional array in numpy_read");
+
+shape.resize(dims.size());
+for(int i=0; i<dims.size(); ++i)
+	{
+        shape[i].number=1;
+        shape[i].symmetry=0;
+#ifndef LA_TENSOR_ZERO_OFFSET
+        shape[i].offset=0;
+#endif
+        shape[i].range=dims[i];
+	}
+
+data.resize(calcsize());
+calcrank(); 
+canonicalize_shape();
+if(sizeof(T)*data.size()!= read(fd,&data[0],sizeof(T)*data.size())) laerror("cannot read 4 in numpy_read");
+if(close(fd)) laerror("cannot close in numpy_read");
+}
+
 
 
 template class Tensor<double>;
diff --git a/tensor.h b/tensor.h
index 2a2fc9c..65823fe 100644
--- a/tensor.h
+++ b/tensor.h
@@ -279,6 +279,10 @@ public:
 	//subtensor - todo: more general versions
 	Tensor subtensor1(int i) const; //for one value of rightmost index
 
+	//numpy interface
+	void numpy_write(const char *name, const char *descr="<f8") const;
+	void numpy_read(const char *name);	
+
 	//conversions from/to matrix and vector
 	explicit Tensor(const NRVec<T> &x);
 	explicit Tensor(const NRMat<T> &x, bool flat=false);