diff --git a/bin/lib/onboard.js b/bin/lib/onboard.js
index 02aa47904d4..a74950fdf5f 100644
--- a/bin/lib/onboard.js
+++ b/bin/lib/onboard.js
@@ -2748,7 +2748,20 @@ async function setupNim(gpu) {
         console.log("  Installing Ollama via Homebrew...");
         run("brew install ollama", { ignoreError: true });
         console.log("  Starting Ollama...");
-        run("OLLAMA_HOST=0.0.0.0:11434 ollama serve > /dev/null 2>&1 &", { ignoreError: true });
+        // On macOS, Docker Desktop routes host-gateway through the VM so
+        // 127.0.0.1 is reachable from containers — bind to localhost to
+        // avoid exposing Ollama to the LAN (CWE-668, NVBUG 6014821).
+        // On Linux, containers access the host via the Docker bridge IP
+        // so 0.0.0.0 is required for reachability.
+        // On WSL2, the default binding works without override.
+        let ollamaEnv = "";
+        if (!isWsl()) {
+          ollamaEnv =
+            process.platform === "darwin"
+              ? "OLLAMA_HOST=127.0.0.1:11434 "
+              : "OLLAMA_HOST=0.0.0.0:11434 ";
+        }
+        run(`${ollamaEnv}ollama serve > /dev/null 2>&1 &`, { ignoreError: true });
         sleep(2);
         console.log("  ✓ Using Ollama on localhost:11434");
         provider = "ollama-local";
diff --git a/spark-install.md b/spark-install.md
index 94cf265bfa5..ef2254d5ea2 100644
--- a/spark-install.md
+++ b/spark-install.md
@@ -86,11 +86,13 @@ ollama run nemotron-3-super:120b
 # type /bye to exit
 ```
 
-### 4. Configure Ollama to Listen on All Interfaces
+### 4. Configure Ollama to Listen on All Interfaces (Linux only)
 
-By default Ollama binds to `127.0.0.1`, which is not reachable from inside the sandbox container. Configure it to listen on all interfaces:
+On Linux, Ollama's default `127.0.0.1` binding is not reachable from inside the sandbox container because Docker containers access the host via the bridge IP, not loopback. Configure Ollama to listen on all interfaces:
 
 > **Note:** `OLLAMA_HOST=0.0.0.0` exposes Ollama on your network. If you're not on a trusted LAN, restrict access with host firewall rules (`ufw`, `iptables`, etc.).
+>
+> **macOS users:** Docker Desktop routes `host-gateway` through the VM, so the default `127.0.0.1` binding works — skip this step. The NemoClaw onboarding wizard handles this automatically.
 
 ```bash
 sudo mkdir -p /etc/systemd/system/ollama.service.d
diff --git a/src/lib/local-inference.test.ts b/src/lib/local-inference.test.ts
index 34040c81475..13bbbcfcaec 100644
--- a/src/lib/local-inference.test.ts
+++ b/src/lib/local-inference.test.ts
@@ -29,9 +29,7 @@ describe("local inference helpers", () => {
   });
 
   it("returns the expected base URL for ollama-local", () => {
-    expect(getLocalProviderBaseUrl("ollama-local")).toBe(
-      "http://host.openshell.internal:11434/v1",
-    );
+    expect(getLocalProviderBaseUrl("ollama-local")).toBe("http://host.openshell.internal:11434/v1");
   });
 
   it("returns null for unknown local provider URLs", () => {
@@ -91,7 +89,13 @@ describe("local inference helpers", () => {
     });
     expect(result.ok).toBe(false);
     expect(result.message).toMatch(/host\.openshell\.internal:11434/);
-    expect(result.message).toMatch(/0\.0\.0\.0:11434/);
+    // Platform-aware message: macOS advises restarting Docker Desktop;
+    // Linux advises binding to 0.0.0.0 (CWE-668 / NVBUG 6014821).
+    if (process.platform === "darwin") {
+      expect(result.message).toMatch(/Restart Docker Desktop/);
+    } else {
+      expect(result.message).toMatch(/0\.0\.0\.0:11434/);
+    }
   });
 
   it("returns a clear error when vllm-local is unavailable", () => {
@@ -202,9 +206,10 @@ describe("local inference helpers", () => {
     expect(
       getBootstrapOllamaModelOptions({ totalMemoryMB: LARGE_OLLAMA_MIN_MEMORY_MB - 1 }),
     ).toEqual(["qwen2.5:7b"]);
-    expect(
-      getBootstrapOllamaModelOptions({ totalMemoryMB: LARGE_OLLAMA_MIN_MEMORY_MB }),
-    ).toEqual(["qwen2.5:7b", DEFAULT_OLLAMA_MODEL]);
+    expect(getBootstrapOllamaModelOptions({ totalMemoryMB: LARGE_OLLAMA_MIN_MEMORY_MB })).toEqual([
+      "qwen2.5:7b",
+      DEFAULT_OLLAMA_MODEL,
+    ]);
     expect(getDefaultOllamaModel(() => "", { totalMemoryMB: 16384 })).toBe("qwen2.5:7b");
   });
 
diff --git a/src/lib/local-inference.ts b/src/lib/local-inference.ts
index 9390bb70ebb..37f939f99f3 100644
--- a/src/lib/local-inference.ts
+++ b/src/lib/local-inference.ts
@@ -119,7 +119,9 @@ export function validateLocalProvider(
       return {
         ok: false,
         message:
-          "Local Ollama is responding on localhost, but containers cannot reach http://host.openshell.internal:11434. Ensure Ollama listens on 0.0.0.0:11434 instead of 127.0.0.1 so sandboxes can reach it.",
+          process.platform === "darwin"
+            ? "Local Ollama is responding on localhost, but containers cannot reach http://host.openshell.internal:11434. Restart Docker Desktop and ensure host networking is enabled."
+            : "Local Ollama is responding on localhost, but containers cannot reach http://host.openshell.internal:11434. Ensure Ollama listens on 0.0.0.0:11434 (not 127.0.0.1) so sandboxes can reach it via the Docker bridge.",
       };
     default:
       return {
@@ -207,10 +209,7 @@ export function getOllamaProbeCommand(
   return `curl -sS --max-time ${timeoutSeconds} http://localhost:11434/api/generate -H 'Content-Type: application/json' -d ${shellQuote(payload)} 2>/dev/null`;
 }
 
-export function validateOllamaModel(
-  model: string,
-  runCapture: RunCaptureFn,
-): ValidationResult {
+export function validateOllamaModel(model: string, runCapture: RunCaptureFn): ValidationResult {
   const output = runCapture(getOllamaProbeCommand(model), { ignoreError: true });
   if (!output) {
     return {
diff --git a/test/e2e/test-gpu-e2e.sh b/test/e2e/test-gpu-e2e.sh
index 6d034b964ca..55d32444043 100755
--- a/test/e2e/test-gpu-e2e.sh
+++ b/test/e2e/test-gpu-e2e.sh
@@ -7,7 +7,7 @@
 # Mirrors what a user with a GPU would actually do:
 #   1. Install Ollama binary
 #   2. Run the NemoClaw installer with NEMOCLAW_PROVIDER=ollama
-#   3. Onboard starts Ollama (OLLAMA_HOST=0.0.0.0:11434), pulls model, creates sandbox
+#   3. Onboard starts Ollama (OLLAMA_HOST=0.0.0.0:11434 on Linux for container reachability), pulls model, creates sandbox
 #   4. Verify inference works through the sandbox
 #   5. Destroy + uninstall
 #
@@ -159,10 +159,10 @@ if [ "${NEMOCLAW_ACCEPT_THIRD_PARTY_SOFTWARE:-}" != "1" ]; then
   exit 1
 fi
 
-# Verify port 11434 is free (onboard needs to start Ollama on 0.0.0.0:11434)
+# Verify port 11434 is free (onboard needs to start Ollama on 0.0.0.0:11434 on Linux for container reachability)
 if curl -sf http://localhost:11434/api/tags >/dev/null 2>&1; then
   info "WARNING: Something is already listening on port 11434."
-  info "Onboard may not be able to start Ollama on 0.0.0.0:11434."
+  info "Onboard may not be able to bind Ollama to 0.0.0.0:11434."
   info "On ephemeral runners this should not happen."
   # Don't fail — onboard will detect the running Ollama and use it.
   # The container reachability check in onboard will catch 127.0.0.1 issues.
@@ -188,7 +188,7 @@ else
 fi
 
 # If the Ollama installer started a system service, stop it so onboard
-# can start Ollama with OLLAMA_HOST=0.0.0.0:11434 (required for containers).
+# can start Ollama with OLLAMA_HOST=0.0.0.0:11434 (required for container reachability on Linux).
 # This needs the ollama process to be owned by our user, or systemctl access.
 if curl -sf http://localhost:11434/api/tags >/dev/null 2>&1; then
   info "Ollama service is running — attempting to stop for clean onboard..."