diff --git a/lindi/LindiH5pyFile/LindiH5pyDataset.py b/lindi/LindiH5pyFile/LindiH5pyDataset.py
index 8e12b93..c1423f0 100644
--- a/lindi/LindiH5pyFile/LindiH5pyDataset.py
+++ b/lindi/LindiH5pyFile/LindiH5pyDataset.py
@@ -229,9 +229,9 @@ def _get_item_for_zarr(self, zarr_array: zarr.Array, selection: Any):
                 )
                 return ret
             else:
-                raise TypeError(
-                    f"Compound dataset {self.name} does not support selection with {selection}"
-                )
+                # Integer or slice indexing on a compound dataset - return
+                # rows as numpy structured array (or np.void for scalar index)
+                return self._get_compound_rows(zarr_array, selection)
 
         # We use zarr's slicing, except in the case of a scalar dataset
         if self.ndim == 0:
@@ -243,6 +243,36 @@ def _get_item_for_zarr(self, zarr_array: zarr.Array, selection: Any):
             return zarr_array[:][0]
         return decode_references(zarr_array[selection])
 
+    def _get_compound_rows(self, zarr_array: zarr.Array, selection):
+        """Return rows from a compound dataset as a numpy structured array.
+
+        For integer indexing, returns a single np.void. For slices, returns
+        a numpy structured array with the compound dtype.
+        """
+        assert self._compound_dtype is not None
+        raw = zarr_array[selection]
+        # raw is either a single list (integer index) or list of lists (slice)
+        if isinstance(selection, (int, np.integer)):
+            # Single row - return np.void
+            row = raw
+            tup = tuple(
+                LindiH5pyReference(row[i]['_REFERENCE']) if isinstance(row[i], dict) and '_REFERENCE' in row[i]
+                else row[i]
+                for i in range(len(self._compound_dtype))
+            )
+            return np.void(tup, dtype=self._compound_dtype)
+        else:
+            # Multiple rows - return structured array
+            result = np.empty(len(raw), dtype=self._compound_dtype)
+            for row_idx, row in enumerate(raw):
+                tup = tuple(
+                    LindiH5pyReference(row[i]['_REFERENCE']) if isinstance(row[i], dict) and '_REFERENCE' in row[i]
+                    else row[i]
+                    for i in range(len(self._compound_dtype))
+                )
+                result[row_idx] = tup
+            return result
+
     def _get_external_hdf5_client(self, url: str) -> h5py.File:
         if url not in _external_hdf5_clients:
             if url.startswith("http://") or url.startswith("https://"):
diff --git a/tests/test_numpy_array_conversion.py b/tests/test_numpy_array_conversion.py
index f2afffa..69abb9e 100644
--- a/tests/test_numpy_array_conversion.py
+++ b/tests/test_numpy_array_conversion.py
@@ -74,3 +74,16 @@ def test_numpy_array_conversion_compound():
 
             y_vals = np.asarray(ds['y'][:])
             np.testing.assert_array_equal(y_vals, np.array([2.5, 4.5, 6.5], dtype=np.float64))
+
+            # Test integer indexing - returns np.void
+            row0 = ds[0]
+            assert isinstance(row0, np.void)
+            assert row0['x'] == 1
+            assert row0['y'] == 2.5
+
+            # Test slice indexing - returns structured array
+            rows = ds[0:2]
+            assert rows.dtype == compound_dtype
+            assert len(rows) == 2
+            np.testing.assert_array_equal(rows['x'], np.array([1, 3], dtype=np.int32))
+            np.testing.assert_array_equal(rows['y'], np.array([2.5, 4.5], dtype=np.float64))