diff --git a/docker-compose.yml b/docker-compose.yml
index ae75ea18d..ec29de645 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -84,13 +84,10 @@ services:
     # cuda / continuum-core-vulkan overlays) it's the actual ceiling.
     mem_limit: ${CONTINUUM_CORE_MEM:-16g}
     working_dir: /app
-    # depends_on does NOT include postgres — postgres is opt-in (profile),
-    # and by default continuum-core uses SQLite where no startup ordering
-    # matters. When users enable the postgres profile and set DATABASE_URL,
-    # Rust's PostgresAdapter (deadpool pool) retries connection on startup.
-    depends_on:
-      livekit-bridge:
-        condition: service_healthy
+    # No depends_on for services behind profiles (postgres, livekit-bridge).
+    # Core starts independently; connections to optional services (postgres
+    # pool, livekit bridge socket) retry on demand. Text chat works without
+    # any profile active — voice/video requires `--profile live`.
     volumes:
       - voice-models:/app/models:ro
       # Mount the ENTIRE ~/.continuum directory R/W. The Rust core reads config,
@@ -130,15 +127,18 @@ services:
   # ── LiveKit Bridge (Rust — WebRTC transport adapter) ──────
   # Links webrtc-sys but NOT ort. Separate process eliminates
   # the protobuf symbol conflict that deadlocked continuum-core.
+  #
+  # Behind `live` profile: voice/video chat is opt-in. Text chat (the
+  # default first-chat experience) doesn't need LiveKit at all. This
+  # saves ~300MB RAM + 3 ports (7880-7882) for Carl's first run.
+  # Enable with: docker compose --profile live up
   livekit-bridge:
+    profiles: [live]
     build:
       context: ./src/workers
       dockerfile: ../../docker/livekit-bridge.Dockerfile
     image: ghcr.io/cambriantech/continuum-livekit-bridge:${CONTINUUM_IMAGE_TAG:-latest}
     restart: unless-stopped
-    # WebRTC encode/decode buffers + multi-stream. Scales with host RAM —
-    # install.sh sets LIVEKIT_BRIDGE_MEM to max(2, host_gb/8). Default 2g
-    # for manual docker compose users; install.sh writes the calculated one.
     mem_limit: ${LIVEKIT_BRIDGE_MEM:-2g}
     depends_on:
       - livekit
@@ -184,7 +184,12 @@ services:
       - NODE_ENV=production
       - JTAG_SKIP_HTTP=1
       - JTAG_NO_TLS=1
-      - LIVEKIT_URL=${LIVEKIT_BROWSER_URL:-ws://livekit:7880}
+      # Browser connects to LiveKit via host-mapped port, not Docker DNS.
+      # 'ws://livekit:7880' only resolves inside the Docker network;
+      # the browser runs on the host where 'livekit' doesn't resolve.
+      # localhost:7880 works because livekit binds that port to the host.
+      # Grid mode overrides via LIVEKIT_BROWSER_URL=ws://tailscale:7880.
+      - LIVEKIT_URL=${LIVEKIT_BROWSER_URL:-ws://localhost:7880}
 
   # ── Widget Server (Vite) ──────────────────────────────────
   widget-server:
@@ -208,10 +213,11 @@ services:
       - JTAG_WS_PROXY_PORT=9001
 
   # ── LiveKit (WebRTC) — local mode ───────────────────────────
-  # Dev server for local development. Always starts.
-  # In grid mode, set LIVEKIT_HOST_PORT=0 in .env to avoid port conflict with tailscale.
-  # (LiveKit still runs but on unmapped ports — harmless, ~50MB RAM.)
+  # Dev server for voice/video. Behind `live` profile — text chat doesn't
+  # need it. In grid mode, set LIVEKIT_HOST_PORT=0 to avoid port conflict.
+  # Enable with: docker compose --profile live up
   livekit:
+    profiles: [live]
     image: livekit/livekit-server:latest
     restart: unless-stopped
     mem_limit: 256m