diff --git a/.evergreen/scripts/setup_tests.py b/.evergreen/scripts/setup_tests.py
index d1ff37242..f3d86973b 100644
--- a/.evergreen/scripts/setup_tests.py
+++ b/.evergreen/scripts/setup_tests.py
@@ -431,6 +431,9 @@ def handle_test_env() -> None:
         # We do not want the default client_context to be initialized.
         write_env("DISABLE_CONTEXT")
 
+    if test_name == "numpy":
+        UV_ARGS.append("--with numpy")
+
     if test_name == "perf":
         data_dir = ROOT / "specifications/source/benchmarking/data"
         if not data_dir.exists():
diff --git a/.evergreen/scripts/utils.py b/.evergreen/scripts/utils.py
index 2bc9c720d..02238645c 100644
--- a/.evergreen/scripts/utils.py
+++ b/.evergreen/scripts/utils.py
@@ -44,6 +44,7 @@ TEST_SUITE_MAP = {
     "mockupdb": "mockupdb",
     "ocsp": "ocsp",
     "perf": "perf",
+    "numpy": "",
 }
 
 # Tests that require a sub test suite.
diff --git a/justfile b/justfile
index be4b69a9b..82b1ac91d 100644
--- a/justfile
+++ b/justfile
@@ -60,12 +60,9 @@ test *args="-v --durations=5 --maxfail=10": && resync
     uv run --extra test python -m pytest {{args}}
 
 [group('test')]
-test-numpy: && resync
-    if [ "${COVERAGE:-}" = "1" ]; then \
-        uv run --extra test --group coverage --with numpy coverage run -m pytest test/test_bson.py; \
-    else \
-        uv run --extra test --with numpy python -m pytest test/test_bson.py; \
-    fi
+test-numpy *args="": && resync
+    just setup-tests numpy {{args}}
+    just run-tests test/test_bson.py
 
 [group('test')]
 run-tests *args: && resync