diff --git a/run_alphafold.py b/run_alphafold.py
index b7c6590..0cd8d87 100644
--- a/run_alphafold.py
+++ b/run_alphafold.py
@@ -32,7 +32,7 @@
 import time
 import typing
 from typing import overload
-
+import tarfile
 from absl import app
 from absl import flags
 from alphafold3.common import folding_input
@@ -303,7 +303,12 @@
     ' and is non-empty. Useful to set this to True to run the data pipeline and'
     ' the inference separately, but use the same output directory.',
 )
-
+_COMPRESS_OUTPUT_DIR = flags.DEFINE_bool(
+    'compress_output_dir',
+    False,
+    'If True, compress the entire output directory into a single .tar.gz archive '
+    'after all outputs are written, and remove the uncompressed version.',
+)
 
 def make_model_config(
     *,
@@ -528,6 +533,7 @@ def write_outputs(
     all_inference_results: Sequence[ResultsForSeed],
     output_dir: os.PathLike[str] | str,
     job_name: str,
+    compress_output_dir: bool = False,
 ) -> None:
   """Writes outputs to the specified output directory."""
   ranking_scores = []
@@ -590,6 +596,12 @@ def write_outputs(
       writer.writerow(['seed', 'sample', 'ranking_score'])
       writer.writerows(ranking_scores)
 
+  if compress_output_dir:
+    archive_path = f"{output_dir}.tar.gz"
+    with tarfile.open(archive_path, "w:gz") as tar:
+      tar.add(output_dir, arcname=os.path.basename(output_dir))
+      shutil.rmtree(output_dir)
+      print(f"Compressed outputs to {archive_path} and removed {output_dir}")
 
 def replace_db_dir(path_with_db_dir: str, db_dirs: Sequence[str]) -> str:
   """Replaces the DB_DIR placeholder in a path with the given DB_DIR."""
@@ -618,6 +630,7 @@ def process_fold_input(
     conformer_max_iterations: int | None = None,
     resolve_msa_overlaps: bool = True,
     force_output_dir: bool = False,
+    compress_output_dir: bool = False
 ) -> folding_input.Input:
   ...
 
@@ -633,6 +646,7 @@ def process_fold_input(
     conformer_max_iterations: int | None = None,
     resolve_msa_overlaps: bool = True,
     force_output_dir: bool = False,
+    compress_output_dir: bool = False
 ) -> Sequence[ResultsForSeed]:
   ...
 
@@ -647,6 +661,7 @@ def process_fold_input(
     conformer_max_iterations: int | None = None,
     resolve_msa_overlaps: bool = True,
     force_output_dir: bool = False,
+    compress_output_dir: bool = False
 ) -> folding_input.Input | Sequence[ResultsForSeed]:
   """Runs data pipeline and/or inference on a single fold input.
 
@@ -733,6 +748,7 @@ def process_fold_input(
         all_inference_results=all_inference_results,
         output_dir=output_dir,
         job_name=fold_input.sanitised_name(),
+        compress_output_dir=compress_output_dir
     )
     output = all_inference_results
 
@@ -887,6 +903,7 @@ def main(_):
         conformer_max_iterations=_CONFORMER_MAX_ITERATIONS.value,
         resolve_msa_overlaps=_RESOLVE_MSA_OVERLAPS.value,
         force_output_dir=_FORCE_OUTPUT_DIR.value,
+        compress_output_dir=_COMPRESS_OUTPUT_DIR.value,
     )
     num_fold_inputs += 1