diff --git a/controller/defaults.py b/controller/defaults.py
index 2e469170..ebd32065 100644
--- a/controller/defaults.py
+++ b/controller/defaults.py
@@ -78,6 +78,16 @@ DEFAULT_CONFIG = {
             "use_https": True,
             "description": "Kiwix/Zim server",
         },
+        {
+            "name": "wrolpi-llm",
+            "systemd_name": "wrolpi-llm",
+            "port": 11435,
+            "viewable": False,
+            "use_https": False,
+            "description": "AI assistant (llama-server)",
+            # On-demand: started by the Controller when AI is enabled, stopped by the API when idle.
+            "kind": "persistent",
+        },
         # System services
         {
             "name": "caddy",
diff --git a/controller/lib/docker_services.py b/controller/lib/docker_services.py
index 82459363..e01f76f7 100644
--- a/controller/lib/docker_services.py
+++ b/controller/lib/docker_services.py
@@ -27,7 +27,7 @@ except ImportError:
 CONTAINER_PREFIX = os.environ.get("COMPOSE_PROJECT_NAME", "wrolpi")
 
 # Containers that should not have a "View" button (non-HTTP services)
-NON_VIEWABLE_CONTAINERS = {"redis", "memcached", "samba"}
+NON_VIEWABLE_CONTAINERS = {"redis", "memcached", "samba", "llm"}
 
 # Containers that should use HTTPS
 HTTPS_CONTAINERS = {"web"}
diff --git a/controller/test/test_defaults.py b/controller/test/test_defaults.py
index 960721d3..6c8bbac4 100644
--- a/controller/test/test_defaults.py
+++ b/controller/test/test_defaults.py
@@ -57,7 +57,7 @@ class TestManagedServicesConfig:
     def test_managed_services_count(self):
         """Should have expected number of managed services."""
         services = DEFAULT_CONFIG["managed_services"]
-        assert len(services) == 11
+        assert len(services) == 12
 
     @pytest.mark.parametrize("name,expected", [
         ("wrolpi-api",
@@ -68,6 +68,8 @@ class TestManagedServicesConfig:
         ("wrolpi-app-dev", {"port": 3000, "viewable": False, "show_only_when_running": True}),
         ("caddy", {"port": 443}),
         ("wrolpi-upgrade", {"port": None, "viewable": False, "show_only_when_running": True}),
+        ("wrolpi-llm",
+         {"systemd_name": "wrolpi-llm", "port": 11435, "viewable": False, "kind": "persistent"}),
     ])
     def test_service_config(self, name, expected):
         """Should have service configured with expected values."""
diff --git a/debian-live-config/config/hooks/normal/9999-wrolpi.hook.chroot b/debian-live-config/config/hooks/normal/9999-wrolpi.hook.chroot
index afd3c2e5..84c134f6 100644
--- a/debian-live-config/config/hooks/normal/9999-wrolpi.hook.chroot
+++ b/debian-live-config/config/hooks/normal/9999-wrolpi.hook.chroot
@@ -143,6 +143,9 @@ cp /opt/wrolpi/wrolpi/blobs/* /opt/wrolpi-blobs/
 # systemd.
 /opt/wrolpi/wrolpi/scripts/install_map_tools.sh
 
+# Bake in llama.cpp (local AI) so a fresh install does not compile it on first upgrade.
+/opt/wrolpi/wrolpi/scripts/install_llama_cpp.sh
+
 # Install WROLPi Help.
 /opt/wrolpi/scripts/install_help_service.sh
 
diff --git a/docker-compose.yml b/docker-compose.yml
index e719f961..aab5101c 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -42,6 +42,8 @@ services:
       # Sanic 25.x wants 'spawn' but our code uses 'fork' for multiprocessing.
       # This tells Sanic to accept the existing start method.
       - SANIC_START_METHOD_SET=1
+      # llama-server runs in the llm container (native installs use 127.0.0.1).
+      - LLAMA_URL=http://llm:11435
     volumes:
       - './main.py:/opt/wrolpi/main.py'
       - './wrolpi:/opt/wrolpi/wrolpi'
@@ -147,6 +149,21 @@ services:
       test: [ 'CMD-SHELL', 'curl http://127.0.0.1:80' ]
       interval: 1m
 
+  llm: # llama-server (local AI) - started/stopped by the controller on demand; the API stops it when idle.
+    build:
+      context: .
+      dockerfile: docker/llm/Dockerfile
+    restart: "no"
+    depends_on:
+      controller:
+        condition: service_healthy
+    volumes:
+      - ${MEDIA_DIRECTORY:-./test}:/media/wrolpi
+      - './docker/llm/entrypoint.sh:/entrypoint.sh'
+    healthcheck:
+      test: [ 'CMD-SHELL', 'curl http://127.0.0.1:11435/health' ]
+      interval: 1m
+
   samba: # Samba file sharing - started/stopped by the controller based on configured shares.
     build:
       context: .
diff --git a/docker/llm/Dockerfile b/docker/llm/Dockerfile
new file mode 100644
index 00000000..24ef1f6e
--- /dev/null
+++ b/docker/llm/Dockerfile
@@ -0,0 +1,17 @@
+FROM debian:13
+ENV DOCKER=true
+
+RUN apt-get update && apt-get install -y curl ca-certificates
+
+# Build llama-server from the pinned release (single source of truth with native installs).
+COPY wrolpi/scripts/install_llama_cpp.sh /tmp/install_llama_cpp.sh
+RUN /tmp/install_llama_cpp.sh && rm /tmp/install_llama_cpp.sh \
+    && apt-get remove -y build-essential cmake && apt-get autoremove -y && apt-get clean
+
+# The same startup scripts the native install uses.
+COPY scripts/start_llama_server.sh /opt/wrolpi/scripts/start_llama_server.sh
+COPY wrolpi/scripts/read_config_value.sh /opt/wrolpi/wrolpi/scripts/read_config_value.sh
+
+COPY docker/llm/entrypoint.sh /entrypoint.sh
+
+ENTRYPOINT ["/entrypoint.sh"]
diff --git a/docker/llm/entrypoint.sh b/docker/llm/entrypoint.sh
new file mode 100755
index 00000000..2e19fa40
--- /dev/null
+++ b/docker/llm/entrypoint.sh
@@ -0,0 +1,5 @@
+#! /usr/bin/env bash
+# Bind all interfaces so the api container can reach llama-server on the internal docker network.
+# The container publishes no host ports and is not proxied by Caddy.
+export LLAMA_HOST=0.0.0.0
+exec /opt/wrolpi/scripts/start_llama_server.sh
diff --git a/etc/raspberrypios/wrolpi-llm.service b/etc/raspberrypios/wrolpi-llm.service
new file mode 100644
index 00000000..46208744
--- /dev/null
+++ b/etc/raspberrypios/wrolpi-llm.service
@@ -0,0 +1,17 @@
+[Unit]
+Description=WROLPi AI (llama-server)
+PartOf=wrolpi.target
+
+[Service]
+ExecStart=/opt/wrolpi/scripts/start_llama_server.sh
+Group=wrolpi
+User=wrolpi
+WorkingDirectory=/opt/wrolpi/
+Restart=on-failure
+# The lowest priority of any WROLPi service; inference must never starve the API.
+Nice=15
+
+# On-demand service: deliberately NOT enabled at boot and NOT wanted by wrolpi.target.
+# The Controller starts it when the user enables AI; the API stops it when idle.
+[Install]
+WantedBy=default.target
diff --git a/modules/ai/__init__.py b/modules/ai/__init__.py
index a34acee8..0fbb8d43 100644
--- a/modules/ai/__init__.py
+++ b/modules/ai/__init__.py
@@ -1,2 +1,3 @@
 from modules.ai.api import ai_bp  # noqa
 from modules.ai.downloader import ModelDownloader  # noqa  Instantiation registers the downloader.
+from modules.ai import service  # noqa  Import registers the idle-unload perpetual worker.
diff --git a/modules/ai/controller_client.py b/modules/ai/controller_client.py
index bef142e3..383451d4 100644
--- a/modules/ai/controller_client.py
+++ b/modules/ai/controller_client.py
@@ -8,7 +8,7 @@ import os
 from typing import Tuple
 from urllib.parse import urlencode
 
-from wrolpi.common import aiohttp_get, logger
+from wrolpi.common import aiohttp_get, aiohttp_post, logger
 from wrolpi.vars import DOCKERIZED
 
 logger = logger.getChild(__name__)
@@ -25,3 +25,13 @@ async def controller_get(path: str, params: dict = None, timeout: int = 10) -> T
             url = f'{url}?{urlencode(params)}'
     async with aiohttp_get(url, timeout=timeout) as response:
         return response.status, await response.json()
+
+
+async def controller_post(path: str, json_: dict = None, timeout: int = 30) -> Tuple[int, dict]:
+    """POST a Controller endpoint, returning (status, JSON body).  Raises on connection errors."""
+    async with aiohttp_post(f'{CONTROLLER_URL}{path}', json_=json_ or {}, timeout=timeout) as response:
+        try:
+            body = await response.json()
+        except Exception:
+            body = {}
+        return response.status, body
diff --git a/modules/ai/service.py b/modules/ai/service.py
new file mode 100644
index 00000000..2299ee0d
--- /dev/null
+++ b/modules/ai/service.py
@@ -0,0 +1,112 @@
+"""llama-server lifecycle: on-demand start and idle unload.
+
+The API owns the policy (this module); the Controller executes the mechanism through its
+existing service start/stop endpoints — it runs as root natively and manages the `llm`
+container in docker, so no sudoers changes are needed.
+
+Idle unload is mandatory on small devices: a swapped-to-death Pi is worse than a slow first
+token.  The chat handler records activity; the perpetual worker stops llama-server after the
+configured idle minutes."""
+import asyncio
+import os
+import time
+
+from modules.ai.config import get_ai_config
+from modules.ai.controller_client import controller_post
+from wrolpi.api_utils import api_app, perpetual_signal
+from wrolpi.common import aiohttp_get, logger
+from wrolpi.vars import DOCKERIZED
+
+logger = logger.getChild(__name__)
+
+LLAMA_URL = os.environ.get('LLAMA_URL', 'http://127.0.0.1:11435')
+# The Controller's name for the service: the systemd unit natively, the compose service in docker.
+LLM_SERVICE_NAME = 'llm' if DOCKERIZED else 'wrolpi-llm'
+
+# How long to wait for llama-server to answer /health after a cold start (model load included).
+STARTUP_TIMEOUT_SECONDS = 90
+
+
+def record_llm_activity():
+    """Note that the LLM was just used; postpones the idle unload."""
+    try:
+        api_app.shared_ctx.ai_llm_last_request.value = time.time()
+    except AttributeError:
+        # Shared contexts are absent in some unit tests.
+        pass
+
+
+def get_last_llm_activity() -> float:
+    try:
+        return api_app.shared_ctx.ai_llm_last_request.value
+    except AttributeError:
+        return 0.0
+
+
+async def llama_healthy() -> bool:
+    try:
+        async with aiohttp_get(f'{LLAMA_URL}/health', timeout=5) as response:
+            return response.status == 200
+    except Exception:
+        return False
+
+
+async def start_llm_service() -> bool:
+    status, body = await controller_post(f'/api/services/{LLM_SERVICE_NAME}/start')
+    if status != 200:
+        logger.error(f'Controller failed to start {LLM_SERVICE_NAME}: HTTP {status} {body}')
+        return False
+    return True
+
+
+async def stop_llm_service() -> bool:
+    status, body = await controller_post(f'/api/services/{LLM_SERVICE_NAME}/stop')
+    if status != 200:
+        logger.error(f'Controller failed to stop {LLM_SERVICE_NAME}: HTTP {status} {body}')
+        return False
+    return True
+
+
+async def ensure_llama_running(timeout: int = STARTUP_TIMEOUT_SECONDS) -> bool:
+    """Start llama-server on demand and wait for it to answer /health.
+
+    Returns True when llama-server is up.  Cold-start latency (service start + model load) is
+    expected and acceptable."""
+    if await llama_healthy():
+        record_llm_activity()
+        return True
+
+    logger.info(f'Asking the Controller to start {LLM_SERVICE_NAME}')
+    if not await start_llm_service():
+        return False
+
+    started = time.time()
+    while time.time() - started < timeout:
+        if await llama_healthy():
+            record_llm_activity()
+            return True
+        await asyncio.sleep(1)
+    logger.error(f'llama-server did not become healthy within {timeout}s')
+    return False
+
+
+@perpetual_signal(sleep=60)
+async def perpetual_llm_idle_unload():
+    """Stop llama-server after the configured idle minutes.  Only acts after it was used."""
+    try:
+        last_activity = get_last_llm_activity()
+        if not last_activity:
+            return
+        config = get_ai_config()
+        idle_seconds = config.idle_unload_minutes * 60
+        if time.time() - last_activity < idle_seconds:
+            return
+        if not await llama_healthy():
+            # Already stopped (or crashed); nothing to unload.
+            api_app.shared_ctx.ai_llm_last_request.value = 0.0
+            return
+        logger.info(f'Stopping {LLM_SERVICE_NAME} after {config.idle_unload_minutes} idle minutes')
+        if await stop_llm_service():
+            api_app.shared_ctx.ai_llm_last_request.value = 0.0
+    except Exception as e:
+        logger.error('LLM idle unload worker failed', exc_info=e)
diff --git a/modules/ai/test/test_service.py b/modules/ai/test/test_service.py
new file mode 100644
index 00000000..28d74bf5
--- /dev/null
+++ b/modules/ai/test/test_service.py
@@ -0,0 +1,78 @@
+"""Tests for the llama-server lifecycle: on-demand start and idle unload."""
+import time
+from unittest import mock
+
+import pytest
+
+from modules.ai import service
+
+
+@pytest.mark.asyncio
+async def test_ensure_llama_running_already_healthy(async_client):
+    """A healthy llama-server is used as-is and activity is recorded."""
+    with mock.patch('modules.ai.service.llama_healthy', return_value=True), \
+            mock.patch('modules.ai.service.start_llm_service') as start:
+        assert await service.ensure_llama_running() is True
+    start.assert_not_called()
+    assert service.get_last_llm_activity() > 0
+
+
+@pytest.mark.asyncio
+async def test_ensure_llama_running_cold_start(async_client):
+    """An unhealthy llama-server is started via the Controller and polled until healthy."""
+    health = iter([False, False, True])
+
+    async def fake_healthy():
+        return next(health)
+
+    with mock.patch('modules.ai.service.llama_healthy', side_effect=fake_healthy), \
+            mock.patch('modules.ai.service.start_llm_service', return_value=True) as start, \
+            mock.patch('modules.ai.service.asyncio.sleep'):
+        assert await service.ensure_llama_running(timeout=10) is True
+    start.assert_called_once()
+
+
+@pytest.mark.asyncio
+async def test_ensure_llama_running_start_fails(async_client):
+    """A Controller start failure is reported as False, not an exception."""
+    with mock.patch('modules.ai.service.llama_healthy', return_value=False), \
+            mock.patch('modules.ai.service.start_llm_service', return_value=False):
+        assert await service.ensure_llama_running(timeout=1) is False
+
+
+@pytest.mark.asyncio
+async def test_idle_unload(async_client, test_ai_config):
+    """llama-server is stopped only after the configured idle minutes."""
+    from wrolpi.api_utils import api_app
+    from modules.ai.config import get_ai_config
+
+    get_ai_config().idle_unload_minutes = 10
+
+    # Never used: nothing happens.
+    api_app.shared_ctx.ai_llm_last_request.value = 0.0
+    with mock.patch('modules.ai.service.stop_llm_service') as stop:
+        await service.perpetual_llm_idle_unload()
+    stop.assert_not_called()
+
+    # Recently used: not stopped.
+    api_app.shared_ctx.ai_llm_last_request.value = time.time()
+    with mock.patch('modules.ai.service.stop_llm_service') as stop, \
+            mock.patch('modules.ai.service.llama_healthy', return_value=True):
+        await service.perpetual_llm_idle_unload()
+    stop.assert_not_called()
+
+    # Idle past the threshold: stopped, and the activity marker resets.
+    api_app.shared_ctx.ai_llm_last_request.value = time.time() - 11 * 60
+    with mock.patch('modules.ai.service.stop_llm_service', return_value=True) as stop, \
+            mock.patch('modules.ai.service.llama_healthy', return_value=True):
+        await service.perpetual_llm_idle_unload()
+    stop.assert_called_once()
+    assert api_app.shared_ctx.ai_llm_last_request.value == 0.0
+
+    # Idle but already stopped: the marker resets without a stop call.
+    api_app.shared_ctx.ai_llm_last_request.value = time.time() - 11 * 60
+    with mock.patch('modules.ai.service.stop_llm_service') as stop, \
+            mock.patch('modules.ai.service.llama_healthy', return_value=False):
+        await service.perpetual_llm_idle_unload()
+    stop.assert_not_called()
+    assert api_app.shared_ctx.ai_llm_last_request.value == 0.0
diff --git a/pi-gen/stage2/04-wrolpi/03-run-chroot.sh b/pi-gen/stage2/04-wrolpi/03-run-chroot.sh
index 9affc7f8..97851f89 100755
--- a/pi-gen/stage2/04-wrolpi/03-run-chroot.sh
+++ b/pi-gen/stage2/04-wrolpi/03-run-chroot.sh
@@ -45,6 +45,9 @@ git config --global --add safe.directory /opt/wrolpi
 # CPU-heavy step on weak hardware.
 /opt/wrolpi/wrolpi/scripts/install_map_tools.sh
 
+# Bake in llama.cpp (local AI) so a fresh Pi does not compile it on first upgrade.
+/opt/wrolpi/wrolpi/scripts/install_llama_cpp.sh
+
 # Install Python requirements.  Try multiple times because pypi may stop responding.
 python3 -m venv /opt/wrolpi/venv
 /opt/wrolpi/venv/bin/pip3 install -r /opt/wrolpi/requirements.txt ||
diff --git a/scripts/start_llama_server.sh b/scripts/start_llama_server.sh
new file mode 100755
index 00000000..b1e786ba
--- /dev/null
+++ b/scripts/start_llama_server.sh
@@ -0,0 +1,54 @@
+#! /usr/bin/env bash
+# Start llama-server (WROLPi's local AI inference runtime) with the model selected in ai.yaml.
+#
+# AI is opt-in: this refuses to start unless ai.yaml enables it and the active model file exists
+# in the media directory (which also covers the unmounted-media case).  Started/stopped on demand
+# by the Controller; the API's idle-unload worker stops it after inactivity.
+set -u
+
+MEDIA_DIRECTORY="${MEDIA_DIRECTORY:-/media/wrolpi}"
+PROJECT_DIR="${PROJECT_DIR:-/opt/wrolpi}"
+AI_CONFIG="${MEDIA_DIRECTORY}/config/ai.yaml"
+# Docker overrides the host so the api container can reach this container.
+LLAMA_HOST="${LLAMA_HOST:-127.0.0.1}"
+LLAMA_PORT="${LLAMA_PORT:-11435}"
+
+read_ai_config() {
+  MEDIA_DIRECTORY="${MEDIA_DIRECTORY}" WROLPI_CONFIG_FILE="${AI_CONFIG}" \
+    "${PROJECT_DIR}/wrolpi/scripts/read_config_value.sh" "$@"
+}
+
+[ ! -d "${MEDIA_DIRECTORY}" ] && echo "Cannot start llama-server without media directory mounted" && exit 1
+[ ! -f "${AI_CONFIG}" ] && echo "Cannot start llama-server without ${AI_CONFIG}" && exit 1
+
+ENABLED="$(read_ai_config enabled false)"
+if [ "${ENABLED}" != "true" ] && [ "${ENABLED}" != "True" ]; then
+  echo "AI is not enabled in ${AI_CONFIG}; refusing to start"
+  exit 1
+fi
+
+ACTIVE_MODEL="$(read_ai_config active_model)"
+[ -z "${ACTIVE_MODEL}" ] && echo "No active_model in ${AI_CONFIG}; refusing to start" && exit 1
+# Defense in depth: the model must be a bare file name inside the models directory.
+case "${ACTIVE_MODEL}" in
+  */*|*..*)
+    echo "Ignoring unsafe active_model '${ACTIVE_MODEL}'" >&2
+    exit 1
+    ;;
+esac
+MODEL="${MEDIA_DIRECTORY}/ai/models/${ACTIVE_MODEL}"
+[ ! -f "${MODEL}" ] && echo "Model file does not exist: ${MODEL}" && exit 1
+
+CONTEXT_SIZE="$(read_ai_config context_size 8192)"
+case "${CONTEXT_SIZE}" in
+  ''|*[!0-9]*) CONTEXT_SIZE=8192 ;;
+esac
+
+# --jinja enables tool-call chat templates; mmap is llama.cpp's default so cold RAM cost is low.
+# Localhost only on native installs -- never exposed through Caddy.
+exec llama-server \
+  --host "${LLAMA_HOST}" \
+  --port "${LLAMA_PORT}" \
+  --jinja \
+  -c "${CONTEXT_SIZE}" \
+  -m "${MODEL}"
diff --git a/scripts/upgrade.sh b/scripts/upgrade.sh
index 382e4a80..cbfde41e 100755
--- a/scripts/upgrade.sh
+++ b/scripts/upgrade.sh
@@ -258,6 +258,10 @@ rm -f /home/wrolpi/.pgpass
 # Migrate from old Leaflet/renderd/Apache2 map stack to MapLibre GL JS + PMTiles.
 /opt/wrolpi/scripts/install_protomaps.sh || echo "Protomaps migration failed."
 
+# Install llama.cpp for the local AI assistant (version-guarded no-op once installed;
+# the first build takes several minutes on a Raspberry Pi).
+/opt/wrolpi/wrolpi/scripts/install_llama_cpp.sh || echo "llama.cpp install failed, continuing..."
+
 # Download map fonts blob if missing.
 MAP_FONTS_BLOB=/opt/wrolpi-blobs/map-fonts.tar.gz
 if [[ ! -f ${MAP_FONTS_BLOB} || ! -s ${MAP_FONTS_BLOB} ]]; then
diff --git a/wrolpi/contexts.py b/wrolpi/contexts.py
index b4fe473b..4095f0bc 100644
--- a/wrolpi/contexts.py
+++ b/wrolpi/contexts.py
@@ -59,6 +59,9 @@ def attach_shared_contexts(app: Sanic):
     app.shared_ctx.download_manager_disabled.set()
     app.shared_ctx.download_manager_stopped.set()
 
+    # Monotonic-ish timestamp of the last chat/LLM request; the AI idle-unload worker reads it.
+    app.shared_ctx.ai_llm_last_request = multiprocessing.Value(ctypes.c_double, 0.0)
+
     app.shared_ctx.single_tasks_started = multiprocessing.Event()
     app.shared_ctx.flags_initialized = multiprocessing.Event()
     app.shared_ctx.perpetual_tasks_started = multiprocessing.Event()
diff --git a/wrolpi/scripts/install_llama_cpp.sh b/wrolpi/scripts/install_llama_cpp.sh
new file mode 100755
index 00000000..c476e3a1
--- /dev/null
+++ b/wrolpi/scripts/install_llama_cpp.sh
@@ -0,0 +1,65 @@
+#!/usr/bin/env bash
+# Install llama.cpp's llama-server, WROLPi's local AI inference runtime.
+#
+# This is the single source of truth, called by every place that provisions a
+# WROLPi so a version bump only has to be made here:
+#   - scripts/upgrade.sh                                  (on-device upgrade)
+#   - pi-gen/stage2/04-wrolpi/03-run-chroot.sh            (Raspberry Pi image build)
+#   - debian-live-config/config/hooks/normal/9999-wrolpi.hook.chroot  (Debian Live ISO build)
+#   - docker/llm/Dockerfile                               (Docker llm container)
+#
+# Built from the SHA-pinned source tarball (llama.cpp ships no generic Linux
+# binaries for ARM).  The build takes several minutes on a Raspberry Pi; the
+# version marker makes re-runs no-ops until the pinned version changes.
+# Run as root: installs into /usr/local/bin and apt-installs build dependencies.
+set -e
+
+LLAMA_CPP_VERSION="v0.3.0"
+# SHA256 of the source tarball -- arch-independent, one value covers every build.
+LLAMA_CPP_SHA256="d94c02d86db22d68692f6bb5b3854763d5091e52142868dc7251995517c666d1"
+
+# llama-server's version output tracks build numbers, not release tags, so the
+# installed release is recorded in a marker file.
+MARKER_DIR=/usr/local/share/wrolpi
+MARKER="${MARKER_DIR}/llama_cpp_version"
+
+install_llama_cpp() {
+  if [ -f "${MARKER}" ] && [ "$(cat "${MARKER}")" = "${LLAMA_CPP_VERSION}" ] \
+      && command -v llama-server &>/dev/null; then
+    echo "llama.cpp ${LLAMA_CPP_VERSION} already installed"
+    return 0
+  fi
+
+  echo "Building llama.cpp ${LLAMA_CPP_VERSION}..."
+  apt-get install -y build-essential cmake
+
+  local build_dir
+  build_dir=$(mktemp -d)
+  curl -fsSL "https://github.com/ggml-org/llama.cpp/archive/refs/tags/${LLAMA_CPP_VERSION}.tar.gz" \
+    -o /tmp/llama_cpp.tar.gz
+  echo "${LLAMA_CPP_SHA256}  /tmp/llama_cpp.tar.gz" | sha256sum -c -
+  tar -xz -C "${build_dir}" --strip-components=1 -f /tmp/llama_cpp.tar.gz
+  rm -f /tmp/llama_cpp.tar.gz
+
+  # Static-ish single binary; GGML_NATIVE=OFF so an image built in a chroot/QEMU
+  # does not bake in the build host's CPU features.
+  (cd "${build_dir}" &&
+    cmake -B build \
+      -DCMAKE_BUILD_TYPE=Release \
+      -DBUILD_SHARED_LIBS=OFF \
+      -DGGML_NATIVE=OFF \
+      -DLLAMA_CURL=OFF \
+      -DLLAMA_BUILD_TESTS=OFF \
+      -DLLAMA_BUILD_EXAMPLES=OFF \
+      -DLLAMA_BUILD_SERVER=ON &&
+    cmake --build build --target llama-server -j"$(nproc)")
+  install -m 0755 "${build_dir}/build/bin/llama-server" /usr/local/bin/llama-server
+  rm -rf "${build_dir}"
+
+  mkdir -p "${MARKER_DIR}"
+  echo "${LLAMA_CPP_VERSION}" > "${MARKER}"
+  llama-server --version || true
+  echo "llama.cpp ${LLAMA_CPP_VERSION} installed"
+}
+
+install_llama_cpp
diff --git a/wrolpi/scripts/read_config_value.sh b/wrolpi/scripts/read_config_value.sh
index c9f55f7f..e57f0637 100755
--- a/wrolpi/scripts/read_config_value.sh
+++ b/wrolpi/scripts/read_config_value.sh
@@ -20,7 +20,8 @@ KEY="${1:?Usage: read_config_value.sh KEY [DEFAULT]}"
 DEFAULT="${2:-}"
 
 MEDIA_DIRECTORY="${MEDIA_DIRECTORY:-/media/wrolpi}"
-CONFIG="${MEDIA_DIRECTORY}/config/wrolpi.yaml"
+# Another top-level config file (e.g. ai.yaml) may be selected with WROLPI_CONFIG_FILE.
+CONFIG="${WROLPI_CONFIG_FILE:-${MEDIA_DIRECTORY}/config/wrolpi.yaml}"
 
 value=""
 if [ -f "${CONFIG}" ]; then