diff --git a/tests/integration/conftest.py b/tests/integration/conftest.py
index 347003c..5ccd500 100644
--- a/tests/integration/conftest.py
+++ b/tests/integration/conftest.py
@@ -15,16 +15,24 @@
 
 from hotdata import ApiClient, Configuration
 from hotdata.api.connections_api import ConnectionsApi
+from hotdata.api.databases_api import DatabasesApi
 from hotdata.api.datasets_api import DatasetsApi
 from hotdata.api.indexes_api import IndexesApi
 from hotdata.api.saved_queries_api import SavedQueriesApi
 from hotdata.api.secrets_api import SecretsApi
 from hotdata.api.workspaces_api import WorkspacesApi
+from hotdata.models.create_database_request import CreateDatabaseRequest
 
 
 REQUIRED_ENV = ("HOTDATA_SDK_TEST_API_KEY", "HOTDATA_SDK_TEST_WORKSPACE_ID")
 DEFAULT_API_URL = "https://api.hotdata.dev"
 
+# Queries are scoped to a database via the X-Database-Id header. Databases no
+# longer auto-expire, so rather than create one per run (which would leak),
+# tests find-or-create a single stable database by name and reuse it — same
+# pattern as the e2e suite. Name is not unique server-side; we key off it.
+SHARED_DATABASE_NAME = "sdkci-shared"
+
 
 @dataclass(frozen=True)
 class TestEnv:
@@ -79,6 +87,23 @@ def connection_id(env: TestEnv) -> str:
     return env.connection_id
 
 
+@pytest.fixture(scope="session")
+def database_id(api_client: ApiClient) -> str:
+    """Id of the shared `sdkci-shared` database, creating it if absent.
+
+    Queries require an `X-Database-Id` scope; databases persist (no auto-expiry)
+    so we reuse one across runs instead of creating-and-deleting per session.
+    """
+    databases_api = DatabasesApi(api_client)
+    for db in databases_api.list_databases().databases:
+        if db.name == SHARED_DATABASE_NAME:
+            return db.id
+    created = databases_api.create_database(
+        CreateDatabaseRequest(name=SHARED_DATABASE_NAME)
+    )
+    return created.id
+
+
 @pytest.fixture
 def sdkci_name() -> "callable[[str], str]":
     """Returns `sdkci-<scenario>-<uuid8>` so orphans are identifiable.
@@ -109,6 +134,11 @@ def connections_api(api_client: ApiClient) -> ConnectionsApi:
     return ConnectionsApi(api_client)
 
 
+@pytest.fixture
+def databases_api(api_client: ApiClient) -> DatabasesApi:
+    return DatabasesApi(api_client)
+
+
 @pytest.fixture
 def indexes_api(api_client: ApiClient) -> IndexesApi:
     return IndexesApi(api_client)
diff --git a/tests/integration/test_query_async_polling.py b/tests/integration/test_query_async_polling.py
index fbf46df..a4d797b 100644
--- a/tests/integration/test_query_async_polling.py
+++ b/tests/integration/test_query_async_polling.py
@@ -40,13 +40,15 @@ def test_query_async_polling(
     query_api: QueryApi,
     query_runs_api: QueryRunsApi,
     results_api: ResultsApi,
+    database_id: str,
 ) -> None:
     # async=True with a small async_after_ms forces the run to come back as
     # AsyncQueryResponse rather than synchronous. The QueryResponse / async
     # response variants are union-shaped on the client; we treat anything with
     # query_run_id as the start of the polling loop.
     submitted = query_api.query(
-        QueryRequest(var_async=True, async_after_ms=1000, sql="SELECT 1 AS x")
+        QueryRequest(var_async=True, async_after_ms=1000, sql="SELECT 1 AS x"),
+        x_database_id=database_id,
     )
     query_run_id = submitted.query_run_id
     assert query_run_id
diff --git a/tests/integration/test_results_arrow.py b/tests/integration/test_results_arrow.py
index 82d3163..e1a2311 100644
--- a/tests/integration/test_results_arrow.py
+++ b/tests/integration/test_results_arrow.py
@@ -46,13 +46,15 @@ def test_results_arrow(
     query_api: QueryApi,
     query_runs_api: QueryRunsApi,
     results_api: ResultsApi,
+    database_id: str,
 ) -> None:
     submitted = query_api.query(
         QueryRequest(
             var_async=True,
             async_after_ms=1000,
             sql="SELECT 1 AS x, 'hello' AS msg UNION ALL SELECT 2, 'world'",
-        )
+        ),
+        x_database_id=database_id,
     )
     query_run_id = submitted.query_run_id
     assert query_run_id
diff --git a/tests/integration/test_saved_query_versioning.py b/tests/integration/test_saved_query_versioning.py
index 61a875a..aacdbf4 100644
--- a/tests/integration/test_saved_query_versioning.py
+++ b/tests/integration/test_saved_query_versioning.py
@@ -13,7 +13,7 @@
 
 
 def test_saved_query_versioning(
-    saved_queries_api: SavedQueriesApi, sdkci_name
+    saved_queries_api: SavedQueriesApi, sdkci_name, database_id: str
 ) -> None:
     name = sdkci_name("savedq-versioning")
     created_id: str | None = None
@@ -50,7 +50,12 @@ def test_saved_query_versioning(
             f"expected versions 1,2,3 in {sorted(version_numbers)}"
         )
 
-        executed = saved_queries_api.execute_saved_query(created.id)
+        # execute_saved_query runs SQL, so it also needs the database scope.
+        # The endpoint has no typed x_database_id param yet, so set the header
+        # directly via the _headers override.
+        executed = saved_queries_api.execute_saved_query(
+            created.id, _headers={"X-Database-Id": database_id}
+        )
         assert executed.row_count == 1
         assert executed.rows == [[3]]
     finally: