diff --git a/tests/functional_tests/t0_main/core/t7_config_numpy_keys.py b/tests/functional_tests/t0_main/core/t7_config_numpy_keys.py
new file mode 100644
index 00000000000..fbcbe6c0d5d
--- /dev/null
+++ b/tests/functional_tests/t0_main/core/t7_config_numpy_keys.py
@@ -0,0 +1,26 @@
+#!/usr/bin/env python
+"""
+Use numpy in nested dict keys
+
+---
+plugin:
+  - wandb
+assert:
+  - :wandb:runs_len: 1
+  - :wandb:runs[0][config][this]: 2
+  - :wandb:runs[0][config][ok]: {"3": 4}
+  - :wandb:runs[0][config][deeper][again]: {"9": 34}
+  - :wandb:runs[0][config][bad]: {"22": 4}
+  - :wandb:runs[0][exitcode]: 0
+"""
+import numpy as np
+import wandb
+
+wandb.init(
+    config={
+        "this": 2,
+        "ok": {3: 4},
+        "deeper": {"again": {9: 34}},
+        "bad": {np.int64(22): 4},
+    }
+)
diff --git a/tests/unit_tests/test_util.py b/tests/unit_tests/test_util.py
index cc8523144aa..f08530f3002 100644
--- a/tests/unit_tests/test_util.py
+++ b/tests/unit_tests/test_util.py
@@ -688,3 +688,38 @@ def test_resolve_aliases():
 
     aliases = util._resolve_aliases("boom")
     assert aliases == ["boom", "latest"]
+
+
+# Compute recursive dicts for tests
+d_recursive1i = {1: 2, 3: {4: 5}}
+d_recursive1i["_"] = d_recursive1i
+d_recursive2i = {1: 2, 3: {np.int64(44): 5}}
+d_recursive2i["_"] = d_recursive2i
+d_recursive2o = {1: 2, 3: {44: 5}}
+d_recursive2o["_"] = d_recursive2o
+
+
+@pytest.mark.parametrize(
+    "dict_input, dict_output",
+    [
+        ({}, None),
+        ({1: 2}, None),
+        ({1: np.int64(3)}, None),  # dont care about values
+        ({np.int64(3): 4}, {3: 4}),  # top-level
+        ({1: {np.int64(3): 4}}, {1: {3: 4}}),  # nested key
+        ({1: {np.int32(2): 4}}, {1: {2: 4}}),  # nested key
+        (d_recursive1i, None),  # recursive, no numpy
+        (d_recursive2i, d_recursive2o),  # recursive, numpy
+    ],
+)
+def test_sanitize_numpy_keys(dict_input, dict_output):
+    dict_output = dict_output.copy() if dict_output is not None else None
+    output, converted = util._sanitize_numpy_keys(dict_input)
+    assert converted == (dict_output is not None)
+
+    # pytest assert cant handle recursive dicts
+    if dict_output and "_" in dict_output:
+        output.pop("_")
+        dict_output.pop("_")
+
+    assert output == (dict_output or dict_input)
diff --git a/wandb/util.py b/wandb/util.py
index 200dcee5c9f..70e898b1742 100644
--- a/wandb/util.py
+++ b/wandb/util.py
@@ -41,6 +41,7 @@
     Mapping,
     Optional,
     Sequence,
+    Set,
     TextIO,
     Tuple,
     Type,
@@ -575,6 +576,68 @@ def matplotlib_contains_images(obj: Any) -> bool:
     return any(len(ax.images) > 0 for ax in obj.axes)
 
 
+def _numpy_generic_convert(obj: Any) -> Any:
+    obj = obj.item()
+    if isinstance(obj, float) and math.isnan(obj):
+        obj = None
+    elif isinstance(obj, np.generic) and (
+        obj.dtype.kind == "f" or obj.dtype == "bfloat16"
+    ):
+        # obj is a numpy float with precision greater than that of native python float
+        # (i.e., float96 or float128) or it is of custom type such as bfloat16.
+        # in these cases, obj.item() does not return a native
+        # python float (in the first case - to avoid loss of precision,
+        # so we need to explicitly cast this down to a 64bit float)
+        obj = float(obj)
+    return obj
+
+
+def _find_all_matching_keys(
+    d: Dict,
+    match_fn: Callable[[Any], bool],
+    visited: Set[int] = None,
+    key_path: Tuple[Any, ...] = (),
+) -> Generator[Tuple[Tuple[Any, ...], Any], None, None]:
+    """Recursively find all keys that satisfies a match function.
+
+    Args:
+       d: The dict to search.
+       match_fn: The function to determine if the key is a match.
+       visited: Keep track of visited nodes so we dont recurse forever.
+       key_path: Keep track of all the keys to get to the current node.
+    Yields:
+       (key_path, key): The location where the key was found, and the key
+    """
+
+    if visited is None:
+        visited = set()
+    me = id(d)
+    if me not in visited:
+        visited.add(me)
+        for key, value in d.items():
+            if match_fn(key):
+                yield key_path, key
+            if isinstance(value, dict):
+                yield from _find_all_matching_keys(
+                    value,
+                    match_fn,
+                    visited=visited,
+                    key_path=tuple(list(key_path) + [key]),
+                )
+
+
+def _sanitize_numpy_keys(d: Dict) -> Tuple[Dict, bool]:
+    np_keys = list(_find_all_matching_keys(d, lambda k: isinstance(k, np.generic)))
+    if not np_keys:
+        return d, False
+    for key_path, key in np_keys:
+        ptr = d
+        for k in key_path:
+            ptr = ptr[k]
+        ptr[_numpy_generic_convert(key)] = ptr.pop(key)
+    return d, True
+
+
 def json_friendly(  # noqa: C901
     obj: Any,
 ) -> Union[Tuple[Any, bool], Tuple[Union[None, str, float], bool]]:  # noqa: C901
@@ -614,19 +677,7 @@ def json_friendly(  # noqa: C901
         elif obj.size <= 32:
             obj = obj.tolist()
     elif np and isinstance(obj, np.generic):
-        obj = obj.item()
-        if isinstance(obj, float) and math.isnan(obj):
-            obj = None
-        elif isinstance(obj, np.generic) and (
-            obj.dtype.kind == "f" or obj.dtype == "bfloat16"
-        ):
-            # obj is a numpy float with precision greater than that of native python float
-            # (i.e., float96 or float128) or it is of custom type such as bfloat16.
-            # in these cases, obj.item() does not return a native
-            # python float (in the first case - to avoid loss of precision,
-            # so we need to explicitly cast this down to a 64bit float)
-            obj = float(obj)
-
+        obj = _numpy_generic_convert(obj)
     elif isinstance(obj, bytes):
         obj = obj.decode("utf-8")
     elif isinstance(obj, (datetime, date)):
@@ -639,6 +690,8 @@ def json_friendly(  # noqa: C901
         )
     elif isinstance(obj, float) and math.isnan(obj):
         obj = None
+    elif isinstance(obj, dict) and np:
+        obj, converted = _sanitize_numpy_keys(obj)
     else:
         converted = False
     if getsizeof(obj) > VALUE_BYTES_LIMIT: