diff --git a/src/benchflow/providers/litellm_config.py b/src/benchflow/providers/litellm_config.py
index acee4756..19600cfd 100644
--- a/src/benchflow/providers/litellm_config.py
+++ b/src/benchflow/providers/litellm_config.py
@@ -275,7 +275,11 @@ def _route_registered_provider(
         )
 
     if provider_cfg.auth_type == "adc":
-        params = {"model": f"vertex_ai/{bare}"}
+        params = {
+            "model": f"vertex_ai/{bare}",
+            "vertex_project": env.get("GOOGLE_CLOUD_PROJECT", ""),
+            "vertex_location": env.get("GOOGLE_CLOUD_LOCATION", ""),
+        }
         return LiteLLMRoute(
             requested_model=model,
             model_alias=safe_model_alias(model),
diff --git a/src/benchflow/providers/litellm_runtime.py b/src/benchflow/providers/litellm_runtime.py
index fc53f118..e1b22c5a 100644
--- a/src/benchflow/providers/litellm_runtime.py
+++ b/src/benchflow/providers/litellm_runtime.py
@@ -1022,7 +1022,11 @@ async def _upload_runtime_files_to_sandbox(
 
 
 async def _ensure_sandbox_litellm(
-    sandbox: Any, *, venv_dir: str, install_timeout_sec: int = 600
+    sandbox: Any,
+    *,
+    venv_dir: str,
+    extra_requirements: tuple[str, ...] = (),
+    install_timeout_sec: int = 600,
 ) -> str:
     vq = shlex.quote(venv_dir)
     # Prefer uv to bootstrap the venv: many sandbox base images ship a python3
@@ -1030,6 +1034,10 @@ async def _ensure_sandbox_litellm(
     # `python -m venv` and `pip install` fail. uv needs neither (it is the same
     # mechanism the openhands agent install already uses in-sandbox), with a
     # stdlib-venv fallback for images that have a working venv and lack uv.
+    requirements = " ".join(
+        shlex.quote(requirement)
+        for requirement in (LITELLM_VERSION_SPEC, "boto3>=1.40", *extra_requirements)
+    )
     command = f"""
 set -eu
 export PATH="$HOME/.local/bin:$PATH"
@@ -1041,7 +1049,7 @@ async def _ensure_sandbox_litellm(
 fi
 if [ -n "$UV" ]; then
   [ -x {vq}/bin/python ] || "$UV" venv {vq} >/dev/null 2>&1
-  "$UV" pip install --python {vq}/bin/python -q '{LITELLM_VERSION_SPEC}' 'boto3>=1.40'
+  "$UV" pip install --python {vq}/bin/python -q {requirements}
 else
   PY="$(command -v python3 || command -v python)"
   if [ ! -x {vq}/bin/python ]; then
@@ -1051,7 +1059,7 @@ async def _ensure_sandbox_litellm(
     )
   fi
   {vq}/bin/python -m pip install -q --upgrade pip
-  {vq}/bin/python -m pip install -q '{LITELLM_VERSION_SPEC}' 'boto3>=1.40'
+  {vq}/bin/python -m pip install -q {requirements}
 fi
 {vq}/bin/python - <<'PY'
 import litellm
@@ -1167,10 +1175,22 @@ async def _start_sandbox_litellm(
         runtime_dir=runtime_dir,
         config=config,
     )
+    proxy_agent_env = dict(agent_env)
+    extra_requirements: tuple[str, ...] = ()
+    if route.provider_name in {"google-vertex", "anthropic-vertex"}:
+        extra_requirements = ("google-cloud-aiplatform>=1.133.0,<2.0",)
+        adc_json = proxy_agent_env.pop("GOOGLE_APPLICATION_CREDENTIALS_JSON", None)
+        if adc_json:
+            adc_path = f"{runtime_dir}/application_default_credentials.json"
+            await _upload_text(sandbox, adc_json, adc_path, ".json")
+            proxy_agent_env["GOOGLE_APPLICATION_CREDENTIALS"] = adc_path
     python = await _ensure_sandbox_litellm(
-        sandbox, venv_dir=paths["venv"], install_timeout_sec=install_timeout_sec
+        sandbox,
+        venv_dir=paths["venv"],
+        extra_requirements=extra_requirements,
+        install_timeout_sec=install_timeout_sec,
     )
-    env = dict(agent_env)
+    env = proxy_agent_env
     env.update(
         {
             "PYTHONPATH": f"{runtime_dir}:{env.get('PYTHONPATH', '')}",
diff --git a/tests/test_litellm_config.py b/tests/test_litellm_config.py
index 2611eecf..f4447954 100644
--- a/tests/test_litellm_config.py
+++ b/tests/test_litellm_config.py
@@ -37,6 +37,16 @@ def test_bedrock_model_honors_max_thinking_effort_env():
     assert route.litellm_params["reasoning_effort"] == "high"
 
 
+def test_vertex_route_passes_project_and_location_to_litellm():
+    route = resolve_litellm_route(
+        "anthropic-vertex/claude-sonnet-4-6",
+        {"GOOGLE_CLOUD_PROJECT": "skillsbench", "GOOGLE_CLOUD_LOCATION": "global"},
+    )
+
+    assert route.litellm_params["vertex_project"] == "skillsbench"
+    assert route.litellm_params["vertex_location"] == "global"
+
+
 def test_azure_openai_route_uses_resource_and_preview_version():
     route = resolve_litellm_route(
         "azure-foundry-openai/gpt-5.5",
diff --git a/tests/test_litellm_hardening.py b/tests/test_litellm_hardening.py
index c39ee1e8..78c3b148 100644
--- a/tests/test_litellm_hardening.py
+++ b/tests/test_litellm_hardening.py
@@ -405,6 +405,49 @@ async def test_sandbox_litellm_install_uses_configured_setup_timeout():
     assert sandbox.exec_timeouts[install_index] == 901
 
 
+@pytest.mark.asyncio
+async def test_vertex_sandbox_litellm_has_google_runtime_and_local_adc():
+    route = resolve_litellm_route(
+        "anthropic-vertex/claude-sonnet-4-6",
+        {"GOOGLE_CLOUD_PROJECT": "project", "GOOGLE_CLOUD_LOCATION": "region"},
+    )
+    sandbox = _FakeSandbox()
+
+    await runtime_mod._start_sandbox_litellm(
+        sandbox=sandbox,
+        route=route,
+        master_key="sk-master",
+        agent_env={"GOOGLE_APPLICATION_CREDENTIALS_JSON": '{"type":"test"}'},
+        session_id="s",
+        agent_name="claude-agent-acp",
+    )
+
+    install_command = next(
+        command
+        for command in sandbox.exec_calls
+        if "pip install" in command and "litellm" in command
+    )
+    assert "google-cloud-aiplatform>=1.133.0,<2.0" in install_command
+    adc_path = next(
+        path
+        for path in sandbox.uploaded
+        if path.endswith("application_default_credentials.json")
+    )
+    launch_path = next(
+        path for path in sandbox.uploaded if path.endswith("launch_config.json")
+    )
+    assert (
+        json.loads(sandbox.uploaded[launch_path])["env"][
+            "GOOGLE_APPLICATION_CREDENTIALS"
+        ]
+        == adc_path
+    )
+    assert (
+        "GOOGLE_APPLICATION_CREDENTIALS_JSON"
+        not in json.loads(sandbox.uploaded[launch_path])["env"]
+    )
+
+
 @pytest.mark.asyncio
 async def test_sandbox_litellm_stop_imports_usage_and_cleans_up():
     route = resolve_litellm_route(