diff --git a/matlab/beamform_mex.cu b/matlab/beamform_mex.cu
new file mode 100644
index 0000000..d5528ac
--- /dev/null
+++ b/matlab/beamform_mex.cu
@@ -0,0 +1,38 @@
+#include "tcbf.h"  // include the header
+#include "mex.h"
+
+void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[]) {
+  if (nrhs != 7) {
+    mexErrMsgIdAndTxt("beamform:InvalidInput",
+                      "Seven inputs required: a_matrix, rf, bf, pixels, frames, samples, device_id");
+  }
+
+  std::string path_a_matrix = mxArrayToString(prhs[0]);
+  std::string path_rf = mxArrayToString(prhs[1]);
+  std::string path_bf = mxArrayToString(prhs[2]);
+  size_t pixels = mxGetScalar(prhs[3]);
+  size_t frames = mxGetScalar(prhs[4]);
+  size_t samples = mxGetScalar(prhs[5]);
+  unsigned device_id = mxGetScalar(prhs[6]);
+
+  cu::init();
+  cu::Device device(device_id);
+  cu::Context context(CU_CTX_BLOCKING_SYNC, device);
+  cu::Stream stream;
+
+  tcbf::Beamformer beamformer(pixels, frames, samples, device, stream);
+  cu::HostMemory RF(2 * frames * samples);
+  cu::HostMemory BF(2 * pixels * frames * sizeof(unsigned));
+
+  beamformer.read_A_matrix(path_a_matrix);
+  beamformer.read_RF(RF, path_rf);
+  beamformer.process(RF, BF);
+  beamformer.write_BF(BF, path_bf);
+
+
+  int status = 0;
+
+  if (nlhs > 0) {
+    plhs[0] = mxCreateDoubleScalar((double)status);
+  }
+}
diff --git a/matlab/mex_compile_beamform.m b/matlab/mex_compile_beamform.m
new file mode 100644
index 0000000..bd36d9e
--- /dev/null
+++ b/matlab/mex_compile_beamform.m
@@ -0,0 +1,8 @@
+% mex_compile_prepare_a_matrix.m
+% Adjust include and library paths as needed
+mexcuda -v ...
+    -L"/usr/local/cuda/lib64" -lcudart -lcuda ...
+    'NVCCFLAGS=-gencode=arch=compute_89,code=sm_89' ...
+    -ltcbf -lccglib ...
+    beamform_mex.cu
+
diff --git a/matlab/mex_compile_prepare_a_matrix.m b/matlab/mex_compile_prepare_a_matrix.m
index 52de9bb..8b0da29 100644
--- a/matlab/mex_compile_prepare_a_matrix.m
+++ b/matlab/mex_compile_prepare_a_matrix.m
@@ -1,10 +1,8 @@
 % mex_compile_prepare_a_matrix.m
 % Adjust include and library paths as needed
-mexcuda ...
-    -I'../include' ...            % for tcbf.h if needed
-    -I'../src' ...                % for prepare_a_matrix.h
-    -I'/path/to/ccglib/include' ...
-    -I'/path/to/cxxopts/include' ...
-    -L'/path/to/ccglib/lib' -lccglib ...
-    -L'/path/to/cuda/lib64' -lcudart -lcuda ...
+mexcuda -v ...
+    -L"/usr/local/cuda/lib64" -lcudart -lcuda ...
+    'NVCCFLAGS=-gencode=arch=compute_89,code=sm_89' ...
+    -lccglib ...
     prepare_a_matrix_mex.cu ../src/prepare_a_matrix.cu
+