diff --git a/kauldron/data/py/__init__.py b/kauldron/data/py/__init__.py
index c9e117cb..35b93260 100644
--- a/kauldron/data/py/__init__.py
+++ b/kauldron/data/py/__init__.py
@@ -19,6 +19,7 @@
 from kauldron.data.py.base import DataSourceBase
 from kauldron.data.py.data_sources import DataSource
 from kauldron.data.py.data_sources import Tfds
+from kauldron.data.py.data_sources import Json
 from kauldron.data.py.mixtures import Mix
 
 # *****************************************************************************
diff --git a/kauldron/data/py/data_sources.py b/kauldron/data/py/data_sources.py
index 6bdf8681..9653304d 100644
--- a/kauldron/data/py/data_sources.py
+++ b/kauldron/data/py/data_sources.py
@@ -19,6 +19,7 @@
 from collections.abc import Mapping
 import dataclasses
 import functools
+import json
 from typing import Any, Optional
 
 from etils import epath
@@ -55,3 +56,38 @@ def data_source(self) -> grain.RandomAccessDataSource:
         data_dir=self.data_dir,
         decoders=self.decoders,
     )
+
+
+# Should this be part of Grain ?
+@dataclasses.dataclass(frozen=True)
+class JsonDataSource(grain.RandomAccessDataSource):
+  """Json data source.
+
+  Assumes that the json file is a list of examples.
+  """
+
+  path: str
+
+  @functools.cached_property
+  def data(self) -> Mapping[str, Any]:
+    return json.loads(epath.Path(self.path).read_text())
+
+  def __len__(self) -> int:
+    return len(self.data)
+
+  def __getitem__(self, record_key):
+    return self.data[record_key]
+
+
+@dataclasses.dataclass(frozen=True)
+class Json(base.DataSourceBase):
+  """Json pipeline.
+
+  Assumes that the json file is a list of examples.
+  """
+
+  path: str
+
+  @functools.cached_property
+  def data_source(self) -> grain.RandomAccessDataSource:
+    return JsonDataSource(path=self.path)