diff --git a/test/e2e-inference.sh b/test/e2e-inference.sh
new file mode 100644
index 00000000000..f0de29ee1ce
--- /dev/null
+++ b/test/e2e-inference.sh
@@ -0,0 +1,273 @@
+#!/usr/bin/env bash
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# E2E Test: Full NemoClaw stack — gateway → sandbox → provider → inference
+#
+# Prerequisites:
+#   - Docker running
+#   - openshell CLI installed
+#   - Ollama running on localhost:11434 with a model pulled (default: nemotron-mini)
+#
+# Usage:
+#   bash test/e2e-inference.sh                    # defaults
+#   OLLAMA_MODEL=llama3.2 bash test/e2e-inference.sh  # override model
+#
+# NOTE: openshell's Docker build engine interprets .dockerignore differently
+# from standard Docker. /dist incorrectly excludes nemoclaw/dist/. This test
+# temporarily adds an exception during the build. See:
+#   https://github.com/NVIDIA/NemoClaw/issues/126
+
+set -uo pipefail
+
+REPO="$(cd "$(dirname "$0")/.." && pwd)"
+SANDBOX_NAME="e2e-test"
+GATEWAY_NAME="nemoclaw"
+PROVIDER_NAME="ollama-e2e"
+MODEL="${OLLAMA_MODEL:-nemotron-mini}"
+PROVIDER_URL="http://host.docker.internal:11434/v1"
+
+PASS=0
+FAIL=0
+
+pass() { ((PASS++)); printf '\033[1;32m  PASS:\033[0m %s\n' "$1"; }
+fail() { ((FAIL++)); printf '\033[1;31m  FAIL:\033[0m %s\n' "$1"; }
+header() { printf '\n\033[1;36m=== %s ===\033[0m\n' "$1"; }
+info()  { printf '\033[1;34m  [info]\033[0m %s\n' "$1"; }
+
+cleanup() {
+  info "Teardown: removing test sandbox and provider..."
+  openshell sandbox delete "$SANDBOX_NAME" 2>/dev/null || true
+  openshell provider delete "$PROVIDER_NAME" 2>/dev/null || true
+
+  # Restore .dockerignore if we modified it
+  if [ -f "${REPO}/.dockerignore.bak" ]; then
+    mv "${REPO}/.dockerignore.bak" "${REPO}/.dockerignore"
+  fi
+}
+trap cleanup EXIT
+
+# ---------------------------------------------------------------------------
+# Step 0: Preflight
+# ---------------------------------------------------------------------------
+step_preflight() {
+  header "Step 0: Preflight checks"
+
+  if ! curl -sf http://localhost:11434/api/tags &>/dev/null; then
+    fail "Ollama not running on localhost:11434"
+    exit 1
+  fi
+  pass "Ollama is running"
+
+  if ! command -v openshell &>/dev/null; then
+    fail "openshell CLI not found"
+    exit 1
+  fi
+  pass "openshell CLI available ($(openshell --version 2>/dev/null))"
+
+  if ! docker info &>/dev/null; then
+    fail "Docker not running"
+    exit 1
+  fi
+  pass "Docker is running"
+}
+
+# ---------------------------------------------------------------------------
+# Step 1: Clean slate
+# ---------------------------------------------------------------------------
+step_clean() {
+  header "Step 1: Clean slate"
+
+  openshell sandbox delete "$SANDBOX_NAME" 2>/dev/null || true
+  openshell provider delete "$PROVIDER_NAME" 2>/dev/null || true
+  pass "Previous test artifacts cleaned"
+}
+
+# ---------------------------------------------------------------------------
+# Step 2: Gateway
+# ---------------------------------------------------------------------------
+step_gateway() {
+  header "Step 2: Gateway"
+
+  local gw_status
+  gw_status="$(openshell status 2>&1 || true)"
+  if echo "$gw_status" | grep -qi "gateway"; then
+    info "Gateway already running, reusing"
+    pass "Gateway ready"
+    return
+  fi
+
+  info "Starting OpenShell gateway '${GATEWAY_NAME}'..."
+  openshell gateway start --name "$GATEWAY_NAME" 2>&1
+
+  local ok=false
+  for _ in $(seq 1 15); do
+    if openshell status 2>&1 | grep -qi "gateway"; then
+      ok=true
+      break
+    fi
+    sleep 2
+  done
+
+  if $ok; then
+    pass "Gateway '${GATEWAY_NAME}' started"
+  else
+    fail "Gateway failed to start"
+    return 1
+  fi
+}
+
+# ---------------------------------------------------------------------------
+# Step 3: Configure inference provider
+# ---------------------------------------------------------------------------
+step_provider() {
+  header "Step 3: Configure inference provider"
+
+  info "Creating provider '${PROVIDER_NAME}' (Ollama → ${MODEL})..."
+  openshell provider create \
+    --name "$PROVIDER_NAME" \
+    --type openai \
+    --credential "OPENAI_API_KEY=ollama" \
+    --config "OPENAI_BASE_URL=${PROVIDER_URL}" 2>&1
+
+  if openshell provider list 2>&1 | grep -q "$PROVIDER_NAME"; then
+    pass "Provider '${PROVIDER_NAME}' created"
+  else
+    fail "Provider creation failed"
+    return 1
+  fi
+
+  info "Setting inference route → ${PROVIDER_NAME} / ${MODEL}"
+  openshell inference set --provider "$PROVIDER_NAME" --model "$MODEL" 2>&1
+  pass "Inference route configured"
+}
+
+# ---------------------------------------------------------------------------
+# Step 4: Create sandbox
+# ---------------------------------------------------------------------------
+step_sandbox() {
+  header "Step 4: Create sandbox"
+
+  # Workaround: openshell's build engine incorrectly matches /dist against
+  # nemoclaw/dist/. Add an exception during the build.
+  if grep -q '^/dist$' "${REPO}/.dockerignore" 2>/dev/null; then
+    cp "${REPO}/.dockerignore" "${REPO}/.dockerignore.bak"
+    sed -i.tmp '/^\/dist$/a\
+!nemoclaw/dist' "${REPO}/.dockerignore" && rm -f "${REPO}/.dockerignore.tmp"
+    info "Patched .dockerignore (workaround for openshell build context)"
+  fi
+
+  info "Building sandbox image and creating '${SANDBOX_NAME}'..."
+  info "(First run builds the Docker image — this takes a few minutes)"
+
+  local policy_path="${REPO}/nemoclaw-blueprint/policies/openclaw-sandbox.yaml"
+
+  openshell sandbox create \
+    --from "${REPO}/Dockerfile" \
+    --name "$SANDBOX_NAME" \
+    --provider "$PROVIDER_NAME" \
+    --policy "$policy_path" \
+    --no-tty \
+    -- env "CHAT_UI_URL=http://127.0.0.1:18789" nemoclaw-start 2>&1
+
+  if openshell sandbox list 2>&1 | grep -q "$SANDBOX_NAME"; then
+    pass "Sandbox '${SANDBOX_NAME}' created"
+  else
+    fail "Sandbox creation failed"
+    return 1
+  fi
+
+  info "Waiting for sandbox services to initialize..."
+  sleep 5
+}
+
+# ---------------------------------------------------------------------------
+# Step 5: Test inference — "Hello Jensen"
+# ---------------------------------------------------------------------------
+step_inference() {
+  header "Step 5: Test inference — Hello Jensen"
+
+  info "Sending inference request through sandbox via inference.local..."
+
+  local ssh_config_file
+  ssh_config_file="$(mktemp)"
+  openshell sandbox ssh-config "$SANDBOX_NAME" > "$ssh_config_file" 2>/dev/null
+
+  local response
+  response="$(ssh -F "$ssh_config_file" \
+    -o StrictHostKeyChecking=no \
+    -o UserKnownHostsFile=/dev/null \
+    -o ConnectTimeout=10 \
+    -o LogLevel=ERROR \
+    "openshell-${SANDBOX_NAME}" \
+    "curl -s --max-time 60 https://inference.local/v1/chat/completions \
+      -H 'Content-Type: application/json' \
+      -d '{\"model\":\"${MODEL}\",\"messages\":[{\"role\":\"user\",\"content\":\"Who is Jensen Huang? Reply in exactly one sentence.\"}],\"max_tokens\":100}'" \
+  )" || true
+  rm -f "$ssh_config_file"
+
+  echo "  Response: ${response:0:500}"
+
+  if [ -z "$response" ]; then
+    fail "Empty response from inference endpoint"
+    return 1
+  fi
+
+  # Validate: valid JSON with choices[0].message.content
+  local content
+  content="$(echo "$response" | python3 -c "
+import sys, json
+data = json.load(sys.stdin)
+print(data['choices'][0]['message']['content'])
+" 2>/dev/null)" || true
+
+  if [ -z "$content" ]; then
+    fail "Response is not a valid chat completion"
+    echo "  Raw: $response"
+    return 1
+  fi
+  pass "Got valid chat completion response"
+
+  pass "LLM responded with content"
+  echo "  Jensen says: ${content}"
+
+  if echo "$content" | grep -qi "jensen\|nvidia\|ceo\|founder\|huang"; then
+    pass "Response mentions Jensen/NVIDIA (inference is coherent)"
+  else
+    info "Response doesn't explicitly mention Jensen/NVIDIA but inference worked"
+    pass "Inference completed successfully"
+  fi
+}
+
+# ---------------------------------------------------------------------------
+# Main
+# ---------------------------------------------------------------------------
+main() {
+  printf '\033[1m\n========================================\n'
+  printf '  E2E Test: Zero to Hello Jensen\n'
+  printf '========================================\033[0m\n'
+
+  step_preflight
+  step_clean
+  step_gateway
+  step_provider
+  step_sandbox
+  step_inference
+
+  # cleanup runs via EXIT trap
+
+  printf '\n\033[1m--- Results ---\033[0m\n'
+  printf '\033[1;32m  Passed: %d\033[0m\n' "$PASS"
+  printf '\033[1;31m  Failed: %d\033[0m\n' "$FAIL"
+  printf '  Total:  %d\n' "$((PASS + FAIL))"
+
+  if [ "$FAIL" -eq 0 ]; then
+    printf '\n\033[1;32mHello Jensen! All tests passed.\033[0m\n'
+  else
+    printf '\n\033[1;31mSome tests failed.\033[0m\n'
+  fi
+
+  return "$FAIL"
+}
+
+main