diff --git a/src/llamacppctl/docker_ops.py b/src/llamacppctl/docker_ops.py
index 46ccfaf..08e8f44 100644
--- a/src/llamacppctl/docker_ops.py
+++ b/src/llamacppctl/docker_ops.py
@@ -138,6 +138,18 @@ def build_run_command(cfg: ServerConfig) -> list:
         "-e", "HF_HOME=/hf_home",
         "-v", f"{cfg.hf_home}:/hf_home:ro",
         "-p", _port_publish(cfg),
+        # Override the image's baked-in HEALTHCHECK (curl on the llama.cpp
+        # default port 8080), which reports "unhealthy" whenever the server runs
+        # on a different --port. Point it at the actually configured port so
+        # Docker's health status matches reality.
+        "--health-cmd",
+        f"curl -f http://localhost:{cfg.container_port}{cfg.health_endpoint} || exit 1",
+        "--health-interval", "30s",
+        "--health-timeout", "5s",
+        "--health-retries", "3",
+        # Model load (large ctx) can take minutes; don't flap as unhealthy while
+        # it warms up.
+        "--health-start-period", "300s",
         cfg.image,
         "-m", model_in_container,
         "--alias", cfg.model_alias,
diff --git a/tests/test_docker_ops.py b/tests/test_docker_ops.py
index ce21f5c..75a6c33 100644
--- a/tests/test_docker_ops.py
+++ b/tests/test_docker_ops.py
@@ -126,6 +126,16 @@ def test_build_run_command_contains_key_flags():
     assert "--cont-batching" in cmd
 
 
+def test_build_run_command_overrides_healthcheck_to_configured_port():
+    cfg = make_cfg()
+    cmd = build_run_command(cfg)
+    assert "--health-cmd" in cmd
+    # healthcheck must target the configured container_port (8000), not the
+    # image default (8080), and hit the configured health endpoint.
+    assert "curl -f http://localhost:8000/health || exit 1" in cmd
+    assert "--health-start-period" in cmd
+
+
 def test_build_run_command_expose_binds_all_interfaces():
     cfg = make_cfg(expose=True)
     cmd = build_run_command(cfg)