diff --git a/.gitignore b/.gitignore
index 828e9003..c5b85cb9 100644
--- a/.gitignore
+++ b/.gitignore
@@ -47,5 +47,4 @@ coverage.xml
 .cci
 .sfdx
 /src.orig
-/src
 myvenv
\ No newline at end of file
diff --git a/snowfakery/api.py b/snowfakery/api.py
index 7b76ac82..a2946f9d 100644
--- a/snowfakery/api.py
+++ b/snowfakery/api.py
@@ -32,6 +32,8 @@
     "jpg": "snowfakery.output_streams.ImageOutputStream",
     "ps": "snowfakery.output_streams.ImageOutputStream",
     "dot": "snowfakery.output_streams.GraphvizOutputStream",
+    "datapack": "snowfakery.experimental.DataPack",
+    "apex": "snowfakery.experimental.DataPack.ApexDataPack",
     "json": "snowfakery.output_streams.JSONOutputStream",
     "txt": "snowfakery.output_streams.DebugOutputStream",
     "csv": "snowfakery.output_streams.CSVOutputStream",
@@ -251,6 +253,10 @@ def _get_output_streams(dburls, output_files, output_format, output_folder):
 
             if output_stream_cls.uses_folder:
                 output_streams.append(output_stream_cls(output_folder))
+            elif output_folder and str(output_folder) != "." and not output_files:
+                raise exc.DataGenError(
+                    "--output-folder can only be used with --output-file=<something> or --output-format=csv"
+                )
 
         if output_files:
             for f in output_files:
diff --git a/snowfakery/cli.py b/snowfakery/cli.py
index 75b1e67c..7a8326bd 100755
--- a/snowfakery/cli.py
+++ b/snowfakery/cli.py
@@ -276,14 +276,6 @@ def validate_options(
             "Sorry, you need to pick --dburl or --output-file "
             "because they are mutually exclusive."
         )
-    if (
-        output_folder
-        and str(output_folder) != "."
-        and not (output_files or output_format == "csv")
-    ):
-        raise click.ClickException(
-            "--output-folder can only be used with --output-file=<something> or --output-format=csv"
-        )
 
     if target_number and reps:
         raise click.ClickException(
diff --git a/snowfakery/data_generator_runtime.py b/snowfakery/data_generator_runtime.py
index cad83c88..ed988aa2 100644
--- a/snowfakery/data_generator_runtime.py
+++ b/snowfakery/data_generator_runtime.py
@@ -421,6 +421,8 @@ def loop_over_templates_until_finished(self, continuing):
             self.iteration_count += 1
             continuing = True
             self.globals.reset_slots()
+            # let the output stream know that the recipe was finished
+            self.output_stream.complete_recipe()
             self.row_history.reset_locals()
 
     def loop_over_templates_once(self, statement_list, continuing: bool):
diff --git a/snowfakery/experimental/DataPack.py b/snowfakery/experimental/DataPack.py
new file mode 100644
index 00000000..8bd9f4c6
--- /dev/null
+++ b/snowfakery/experimental/DataPack.py
@@ -0,0 +1,210 @@
+# Use this experimental OutputStream like this:
+
+# snowfakery --output-format snowfakery.experimental.DataPack recipe.yml > composite.json
+#
+# Once you have the file you can make it accessible to Salesforce by uploading it
+# to some form of server. E.g. Github gist, Heroku, etc.
+#
+# Then you can use Anon Apex like that in `LoadCompositeAPIData.apex` to load it into
+# any org. e.g.:
+
+# sfdx force:apex:execute -f ./examples/salesforce/LoadCompositeAPIData.apex -u Snowfakery__qa
+# or
+# cci task run execute_anon --path examples/salesforce/LoadCompositeAPIData.apex --org qa
+#
+# Note that Salesforce will complain if the dataset has more than 500 rows.
+
+# TODO: Add tests
+
+import json
+from logging import warning
+from io import StringIO
+import typing as T
+import datetime
+from pathlib import Path
+from tempfile import TemporaryDirectory
+
+from snowfakery.output_streams import FileOutputStream, OutputStream
+
+MAX_BATCH_SIZE = 500  # https://developer.salesforce.com/docs/atlas.en-us.api_rest.meta/api_rest/resources_composite_graph_limits.htm
+
+
+class SalesforceCompositeAPIOutput(FileOutputStream):
+    """Output stream that generates records for Salesforce's Composite API"""
+
+    encoders: T.Mapping[type, T.Callable] = {
+        **FileOutputStream.encoders,
+        datetime.date: str,
+        datetime.datetime: str,
+        bool: bool,
+    }
+    is_text = True
+
+    def __init__(self, file, **kwargs):
+        assert file
+        super().__init__(file, **kwargs)
+        self.rows = []
+
+    def write_single_row(self, tablename: str, row: T.Dict) -> None:
+        row_without_id = row.copy()
+        del row_without_id["id"]
+        _sf_update_key = row_without_id.pop("_sf_update_key", None)
+        if _sf_update_key:
+            method = "PATCH"
+            url = f"/services/data/v50.0/sobjects/{tablename}/{_sf_update_key}/{row_without_id[_sf_update_key]}"
+        else:
+            method = "POST"
+            url = f"/services/data/v50.0/sobjects/{tablename}/"
+
+        values = {
+            "method": method,
+            "referenceId": f"{tablename}_{row['id']}",
+            "url": url,
+            "body": row_without_id,
+        }
+        self.rows.append(values)
+
+    def flatten(
+        self,
+        sourcetable: str,
+        fieldname: str,
+        source_row_dict,
+        target_object_row,
+    ) -> T.Union[str, int]:
+        target_reference = f"{target_object_row._tablename}_{target_object_row.id}"
+        return "@{%s.id}" % target_reference
+
+    def close(self, **kwargs) -> T.Optional[T.Sequence[str]]:
+        # NOTE: Could improve loading performance by breaking graphs up
+        # to allow server-side parallelization, but I'd risk locking issues
+        assert self.rows
+        data = {"graphs": [{"graphId": "graph", "compositeRequest": self.rows}]}
+        self.write(json.dumps(data, indent=2))
+        return super().close()
+
+
+class Folder(OutputStream):
+    uses_folder = True
+
+    def __init__(self, output_folder, **kwargs):
+        super().__init__(None, **kwargs)
+        self.target_path = Path(output_folder)
+        if not Path.exists(self.target_path):
+            Path.mkdir(self.target_path, exist_ok=True)  # pragma: no cover
+        self.recipe_sets = [[]]
+        self.current_batch = []
+        self.filenum = 0
+        self.filenames = []
+
+    def write_row(
+        self, tablename: str, row_with_references: T.Dict, *args, **kwargs
+    ) -> None:
+        self.recipe_sets[-1].append((tablename, row_with_references))
+
+    def write_single_row(self, tablename: str, row: T.Dict, *args, **kwargs) -> None:
+        raise NotImplementedError(
+            "Shouldn't be called. write_row should be called instead"
+        )
+
+    def close(self, **kwargs) -> T.Optional[T.Sequence[str]]:
+        self.flush_sets()
+        self.flush_batch()
+        table_metadata = [{"url": str(filename)} for filename in self.filenames]
+        metadata = {
+            "@context": "http://www.w3.org/ns/csvw",
+            "tables": table_metadata,
+        }
+        metadata_filename = self.target_path / "csvw_metadata.json"
+        with open(metadata_filename, "w") as f:
+            json.dump(metadata, f, indent=2)
+        return [f"Created {self.target_path}"]
+
+    def complete_recipe(self, *args):
+        self.flush_sets()
+        self.recipe_sets.append([])
+
+    def flush_sets(self):
+        while self.recipe_sets:
+            next_set = self.recipe_sets.pop(0)
+            assert len(next_set) <= MAX_BATCH_SIZE
+            if len(self.current_batch) + len(next_set) > MAX_BATCH_SIZE:
+                self.flush_batch()
+            self.current_batch.extend(next_set)
+
+    def flush_batch(self):
+        self.filenum += 1
+        filename = Path(self.target_path) / f"{self.filenum}.composite.json"
+
+        with open(filename, "w") as open_file, SalesforceCompositeAPIOutput(
+            open_file
+        ) as out:
+            self.filenames.append(filename)
+            assert self.current_batch
+            for tablename, row in self.current_batch:
+                out.write_row(tablename, row)
+
+        self.current_batch = []
+
+
+class DataPack(FileOutputStream):
+    def __init__(self, file, **kwargs):
+        super().__init__(file, **kwargs)
+        warning("DataPack is an experimental data format")
+        self.tempdir = TemporaryDirectory()
+        self.folder_os = Folder(self.tempdir.name)
+
+    def write_row(
+        self, tablename: str, row_with_references: T.Dict, *args, **kwargs
+    ) -> None:
+        self.folder_os.write_row(tablename, row_with_references)
+
+    def write_single_row(self, tablename: str, row: T.Dict, *args, **kwargs) -> None:
+        raise NotImplementedError(
+            "Shouldn't be called. write_row should be called instead"
+        )
+
+    def complete_recipe(self, *args):
+        self.folder_os.complete_recipe()
+
+    def close(self):
+        self.folder_os.close()
+        data = self.organize_bundle()
+        self.write(json.dumps(data, indent=2))
+        self.tempdir.cleanup()
+        return super().close()
+
+    def organize_bundle(self):
+        files = Path(self.tempdir.name).glob("*.composite.json")
+        data = [file.read_text() for file in files]
+        assert data
+        return {"datapack_format": 1, "data": data}
+
+
+class ApexDataPack(FileOutputStream):
+    """Wrap in Anon Apex but note that the amount of data you can load
+    this way is very limited due to limitations of the REST API (used by CCI)
+    and SOAP API (used by sfdx)"""
+
+    def __init__(self, file, **kwargs):
+        super().__init__(file, **kwargs)
+        self.data = StringIO()
+        self.datapack = DataPack(self.data)
+
+    def write_row(
+        self, tablename: str, row_with_references: T.Dict, *args, **kwargs
+    ) -> None:
+        self.datapack.write_row(tablename, row_with_references)
+
+    def write_single_row(self, tablename: str, row: T.Dict, *args, **kwargs) -> None:
+        raise NotImplementedError(
+            "Shouldn't be called. write_row should be called instead"
+        )
+
+    def complete_recipe(self, *args):
+        self.datapack.complete_recipe()
+
+    def close(self):
+        self.datapack.close()
+        quoted_data = repr(self.data.getvalue())
+        self.write(f"String json_data = {quoted_data};\n")
+        self.write("LoadCompositeAPIData.loadBundledJsonSet(json_data);\n")
diff --git a/snowfakery/output_streams.py b/snowfakery/output_streams.py
index 5a719c70..1097a1ee 100644
--- a/snowfakery/output_streams.py
+++ b/snowfakery/output_streams.py
@@ -136,6 +136,11 @@ def __enter__(self, *args):
     def __exit__(self, *args):
         self.close()
 
+    def complete_recipe(self, *args):
+        """Let the output stream know that a complete recipe
+        set was generated."""
+        pass
+
 
 class SmartStream:
     """Common code for managing stream/file opening/closing
diff --git a/tests/test_DataPack.py b/tests/test_DataPack.py
new file mode 100644
index 00000000..a306a720
--- /dev/null
+++ b/tests/test_DataPack.py
@@ -0,0 +1,69 @@
+from io import StringIO
+from unittest.mock import patch
+
+from snowfakery.data_generator import generate
+from snowfakery.data_generator_runtime import StoppingCriteria
+from snowfakery.experimental.DataPack import (
+    DataPack,
+    ApexDataPack,
+    SalesforceCompositeAPIOutput,
+)
+import json
+
+## Fill this out when it isn't experimental anymore
+
+
+class TestSalesforceCompositeAPIOutput:
+    @patch("snowfakery.experimental.DataPack.MAX_BATCH_SIZE", 10)
+    def test_composite(self):
+        out = StringIO()
+        output_stream = DataPack(out)
+        with open("examples/basic-salesforce.recipe.yml") as f:
+            generate(
+                f, {}, output_stream, stopping_criteria=StoppingCriteria("Account", 15)
+            )
+        output_stream.close()
+        data = json.loads(out.getvalue())
+        assert data["datapack_format"] == 1
+        assert len(data["data"]) == 8
+        single_payload = json.loads(data["data"][0])
+        print(single_payload)
+        assert single_payload["graphs"][0]["compositeRequest"][0]["method"] == "POST"
+
+    def test_reference(self):
+        out = StringIO()
+        output_stream = SalesforceCompositeAPIOutput(out)
+        with open("examples/basic-salesforce.recipe.yml") as f:
+            generate(f, {}, output_stream)
+        output_stream.close()
+        print(out.getvalue())
+        data = json.loads(out.getvalue())
+        assert (
+            data["graphs"][0]["compositeRequest"][-1]["body"]["AccountId"]
+            == "@{Account_2.id}"
+        )
+
+    @patch("snowfakery.experimental.DataPack.MAX_BATCH_SIZE", 50)
+    def test_composite_upsert(self):
+        out = StringIO()
+        output_stream = DataPack(out)
+        with open("tests/upsert-2.yml") as f:
+            generate(
+                f, {}, output_stream, stopping_criteria=StoppingCriteria("Account", 50)
+            )
+        output_stream.close()
+        data = json.loads(out.getvalue())
+        assert data["datapack_format"] == 1
+        single_payload = json.loads(data["data"][1])
+        assert single_payload["graphs"][0]["compositeRequest"][-1]["method"] == "PATCH"
+
+    def test_apex(self):
+        out = StringIO()
+        output_stream = ApexDataPack(out)
+        with open("examples/basic-salesforce.recipe.yml") as f:
+            generate(
+                f, {}, output_stream, stopping_criteria=StoppingCriteria("Account", 50)
+            )
+        output_stream.close()
+        out = out.getvalue()
+        assert out.startswith("String json_data")
diff --git a/tests/upsert-2.yml b/tests/upsert-2.yml
index bca7a6a8..c48d2adc 100644
--- a/tests/upsert-2.yml
+++ b/tests/upsert-2.yml
@@ -1,6 +1,145 @@
+# Accounts
+
+- object: Account
+  nickname: Bluth Company
+  fields:
+    Name: Bluth Company
+- object: Account
+  nickname: Michael B Company
+  fields:
+    Name: Michael B. Company
+- object: Account
+  nickname: Austero Bluth Company
+  fields:
+    Name: Austero Bluth Company
+- object: Account
+  nickname: Sitwell Enterprises
+  fields:
+    Name: Sitwell Enterprises
+
+# Contacts
+
+- object: Contact
+  update_key: email
+  nickname: GOB
+  fields:
+    firstname: GOB
+    lastname: Bluth
+    email: George.Oscar@bluth.com
+    AccountId:
+      reference: Bluth Company
 - object: Contact
   update_key: email
+  nickname: Michael
+  fields:
+    firstname: Michael
+    lastname: Bluth
+    email: michael@bluth.com
+    AccountId:
+      reference: Michael B Company
+- object: Contact
+  update_key: email
+  nickname: Lucille-2
+  fields:
+    firstname: Lucille
+    lastname: Austero
+    email: Lucille.Austero@bluth.com
+    AccountId:
+      reference: Austero Bluth Company
+- object: Contact
+  update_key: email
+  nickname: Lindsay
+  fields:
+    firstname: Lindsay
+    lastname: Bluth
+    email: lindsay@bluth.com
+    AccountId:
+      reference: Sitwell Enterprises
+- object: Contact
+  update_key: email
+  nickname: anyong
   fields:
     firstname: Hel-Loh
     lastname: Bluth
     email: anyong@bluth.com
+- object: Contact
+  update_key: email
+  nickname: george
+  fields:
+    firstname: George
+    lastname: Bluth
+    email: boss@bluth.com
+- object: Contact
+  update_key: email
+  nickname: buster
+  fields:
+    firstname: Buster
+    lastname: Bluth
+    email: buster@bluth.com
+
+# Opportunities
+
+- object: Opportunity
+  update_key: Name
+  fields:
+    Name: Bluth Company Opp
+    AccountId:
+      reference: Bluth Company
+    # ContactId:
+    #   reference: GOB
+    CloseDate:
+      date_between:
+        start_date: -1y
+        end_date: today
+    StageName:
+      random_choice:
+        - Prospecting
+        - Pledged
+- object: Opportunity
+  update_key: Name
+  fields:
+    Name: Michael B Company Opp
+    AccountId:
+      reference: Michael B Company
+    # ContactId:
+    #   reference: Michael
+    CloseDate:
+      date_between:
+        start_date: -1y
+        end_date: today
+    StageName:
+      random_choice:
+        - Prospecting
+        - Pledged
+- object: Opportunity
+  update_key: Name
+  fields:
+    Name: Austero Bluth Company Opp
+    AccountId:
+      reference: Austero Bluth Company
+    # ContactId:
+    #   reference: Lucille-2
+    CloseDate:
+      date_between:
+        start_date: -1y
+        end_date: today
+    StageName:
+      random_choice:
+        - Prospecting
+        - Pledged
+- object: Opportunity
+  update_key: Name
+  fields:
+    Name: Sitwell Enterprises Opp
+    AccountId:
+      reference: Sitwell Enterprises
+    # ContactId:
+    #   reference: Lindsay
+    CloseDate:
+      date_between:
+        start_date: -1y
+        end_date: today
+    StageName:
+      random_choice:
+        - Prospecting
+        - Pledged