diff --git a/_unittests/ut_onnxrt/test_onnx_tools.py b/_unittests/ut_onnxrt/test_onnx_tools.py
new file mode 100644
index 000000000..ab398679c
--- /dev/null
+++ b/_unittests/ut_onnxrt/test_onnx_tools.py
@@ -0,0 +1,56 @@
+"""
+@brief      test log(time=2s)
+"""
+import unittest
+from logging import getLogger
+import numpy
+from onnx import helper, TensorProto
+from pyquickhelper.pycode import ExtTestCase, ignore_warnings
+from mlprodict.onnxrt import OnnxInference
+from mlprodict.onnxrt.onnx_tools import insert_node
+from mlprodict.onnxrt.ops_cpu._op import RuntimeTypeError
+
+
+class TestOnnxTools(ExtTestCase):
+
+    def setUp(self):
+        logger = getLogger('skl2onnx')
+        logger.disabled = True
+
+    @ignore_warnings(DeprecationWarning)
+    def test_onnx_inference_name_confusion(self):
+        X = helper.make_tensor_value_info(
+            'X', TensorProto.FLOAT, [None, 2])  # pylint: disable=E1101
+        Y = helper.make_tensor_value_info(
+            'Y', TensorProto.FLOAT, [None, 2])  # pylint: disable=E1101
+        Z = helper.make_tensor_value_info(
+            'Z', TensorProto.FLOAT, [None, 2])  # pylint: disable=E1101
+        node_def = helper.make_node('Add', ['X', 'Y'], ['Zt'], name='Zt')
+        node_def2 = helper.make_node('Add', ['X', 'Zt'], ['Z'], name='Z')
+        graph_def = helper.make_graph(
+            [node_def, node_def2], 'test-model', [X, Y], [Z])
+        model_def = helper.make_model(graph_def, producer_name='onnx-example')
+        model_def = insert_node(
+            model_def, node='Z', op_type='Cast', to=TensorProto.INT64,  # pylint: disable=E1101
+            name='castop')
+        self.assertIn('castop', str(model_def))
+
+        oinf = OnnxInference(model_def)
+        X = (numpy.random.randn(4, 2) * 100000).astype(  # pylint: disable=E1101
+            numpy.float32)
+        Y = (numpy.random.randn(4, 2) * 100000).astype(  # pylint: disable=E1101
+            numpy.float32)
+        exp = (X * 2 + Y).astype(numpy.float32)
+        self.assertRaise(lambda: oinf.run({'X': X, 'Y': Y}), RuntimeTypeError)
+
+        model_def = insert_node(
+            model_def, node='Z', op_type='Cast', to=TensorProto.FLOAT,  # pylint: disable=E1101
+            name='castop2')
+        oinf = OnnxInference(model_def)
+        res = oinf.run({'X': X, 'Y': Y})
+        got = res['Z']
+        self.assertEqualArray(exp / 100000, got / 100000, decimal=5)
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/mlprodict/onnxrt/onnx_tools.py b/mlprodict/onnxrt/onnx_tools.py
new file mode 100644
index 000000000..2a7c222f2
--- /dev/null
+++ b/mlprodict/onnxrt/onnx_tools.py
@@ -0,0 +1,104 @@
+"""
+@file
+@brief Functions to manipulate ONNX file.
+"""
+from onnx import helper
+
+
+def find_node_name(model, name):
+    """
+    Finds a node by its name.
+    :param model: onnx graph
+    :param name: node name
+    :return: node pointer
+    """
+    if not hasattr(model, "graph"):
+        raise TypeError(  # pragma: no cover
+            "Parameter model is not an ONNX model but "
+            "{}".format(type(model)))
+    for node in model.graph.node:
+        if node.name == name:
+            return node
+    return None
+
+
+def find_node_input_name(node, name):
+    """
+    Finds a node input by its name.
+    :param node: onnx node
+    :param name: node name
+    :return: input index
+    """
+    for i, inode in enumerate(node.input.node):
+        if inode.name == name:
+            return i
+    return -1
+
+
+def insert_node(model, op_type, node, input_index=0, new_name=None, **attrs):
+    """
+    Inserts a node before one node input.
+    :param model: onnx graph
+    :param op_type:
+    :param node: node or node name
+    :param input_index: input index or input name
+    :param attrs: node attributes
+    :return: updated graph
+    """
+    if isinstance(node, str):
+        inode = find_node_name(model, node)
+    else:
+        inode = node
+    if isinstance(input_index, str):
+        input_index_ = find_node_input_name(node, input_index)
+        if input_index_ == -1:
+            raise RuntimeError(
+                "Unable to find input_index %r in node %r." % (
+                    input_index, node.name))  # pylint: disable=E1120
+        input_index = input_index_
+
+    # guess a new name
+    names = []
+    for n in model.graph.node:
+        names.extend(n.input)
+        names.extend(n.output)
+    names = set(names)
+    if new_name is None:
+        new_name = op_type.lower()
+    root_name = new_name
+    i = 0
+    while new_name in names:
+        new_name = "%s_%d" % (root_name, i)
+        i += 1
+
+    new_node = helper.make_node(
+        op_type, [inode.input[input_index]], [new_name], **attrs)
+    inode.input[input_index] = new_name
+    keep_nodes = list(model.graph.node)
+    keep_nodes.append(new_node)
+
+    graph = helper.make_graph(
+        keep_nodes, model.graph.name, model.graph.input,
+        model.graph.output, model.graph.initializer)
+    onnx_model = helper.make_model(graph)
+    onnx_model.ir_version = model.ir_version
+    onnx_model.producer_name = model.producer_name
+    onnx_model.producer_version = model.producer_version
+    onnx_model.domain = model.domain
+    onnx_model.model_version = model.model_version
+    onnx_model.doc_string = model.doc_string
+    if len(model.metadata_props) > 0:
+        values = {p.key: p.value for p in model.metadata_props}
+        helper.set_model_props(onnx_model, values)
+
+    del onnx_model.opset_import[:]  # pylint: disable=E1101
+    for oimp in model.opset_import:
+        op_set = onnx_model.opset_import.add()  # pylint: disable=E1101
+        op_set.domain = oimp.domain
+        op_set.version = oimp.version
+
+    if len(onnx_model.graph.input) != len(model.graph.input):  # pylint: disable=E1101
+        raise RuntimeError(  # pragma: no cover
+            "Input mismatch {} != {}".format(
+                len(onnx_model.input), len(model.input)))  # pylint: disable=E1101
+    return onnx_model