ODS V3 Pre-Release: Public testing and refinement ahead of the official V3 launch. Turn your PC, Mac, or Linux box into a private AI server.
M1 Offline Mode — Fully Air-Gapped Operation
ODS can keep running after you disconnect it from the network.
Overview
--offline prepares an install for air-gapped operation. The installation
itself still needs the network: it downloads container images, the model and
an embedding model. Once it finishes, chat, voice and local documents keep
working with the machine disconnected.
--offline is an option of the Linux installer (install-core.sh, which the
Windows WSL path also uses). The macOS installer does not have it.
What --offline changes
- Skips the installer's network reachability check and the small bootstrap model, so the full model downloads before launch.
- Clears
BRAVE_API_KEY,OPENAI_API_KEYandANTHROPIC_API_KEYin.env. - Sets
DISABLE_UPDATE_CHECK=true, which stops the dashboard's GitHub release check. - Downloads
nomic-embed-text-v1.5.Q4_K_M.ggufand writes the.offline-modemarker only after that file validates.
What it does not change
- It does not block outbound network access. Air-gapped operation means disconnecting the machine.
- Portal's web search keeps its configured provider and works whenever the
machine is online. Fresh installs use the keyless Parallel API; set
PIXEL_WEB_SEARCH_PROVIDER=searxngbefore installing to search through the bundled SearXNG instead. OFFLINE_MODE,WEB_SEARCH_ENABLED,LOCAL_RAG_ENABLEDandDISABLE_TELEMETRYare written to.env, but no service reads them yet.
Bundled services ship with their own usage telemetry and update checks turned off in every install, not only offline ones: Open WebUI, Qdrant, LiteLLM's cost-map fetch, n8n and the Whisper Hugging Face client.
Installation
# Standard offline install
./install.sh --offline --all
# Offline with specific tier
./install.sh --offline --tier 2 --voice --rag
What Works Offline
| Component | Status | Notes |
|---|---|---|
| llama-server (local LLM) | ✅ | All inference local |
| Open WebUI | ✅ | Local web interface |
| Whisper STT | ✅ | --voice flag; cache the model before disconnecting (below) |
| Kokoro TTS | ✅ | --voice flag |
| Qdrant (RAG) | ✅ | --rag flag |
| Portal (Pixel) | ⚠️ | Default agent on qualified hosts; chat works, web search and fetch need the network |
| Hermes Agent | ✅ | On with Full Stack, --all or --hermes; local LLM only |
| n8n workflows | ⚠️ | Local execution, but many integrations need internet |
Post-Installation
Verify Offline Operation
# Check services are running
ods status
# Test LLM (should work offline)
curl http://localhost:11434/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{"model": "local", "messages": [{"role": "user", "content": "Hello"}]}'
# Test local embeddings/RAG if enabled.
Full Air-Gap Procedure
- Complete installation with
--offlineflag - If voice is enabled, run
ods stt downloadto cache the Whisper model - Verify all services running:
ods status - Test core functionality while online
- Disconnect network (unplug ethernet / disable WiFi)
- Verify services still work
Reconnecting (Optional)
Images and models are pinned, so pulling them again does not update anything. To move to a newer ODS, reconnect, follow How do I update ODS?, then disconnect.
Pre-Downloaded Models
An offline install downloads, while it is still online:
- LLM — Based on your tier selection
- GGUF embeddings —
nomic-embed-text-v1.5.Q4_K_M.gguf(~300MB) - Whisper — If
--voiceis enabled, a later installer phase downloads it; runods stt downloadbefore disconnecting to be sure it is cached - Kokoro TTS image/data — If
--voiceis enabled
Replacing Web Search
Web search needs the internet. Offline, use local documents instead.
Pre-Load a Knowledge Base
# Index local documents into Qdrant
curl -X POST http://localhost:6333/collections/knowledge/points \
-H "Content-Type: application/json" \
-d '{...your documents...}'
Troubleshooting
Memory Search Not Working
Check GGUF embeddings downloaded:
ls -la models/embeddings/
# Should see: nomic-embed-text-v1.5.Q4_K_M.gguf
Services Won't Start Without Internet
All images should be pulled during installation. If one is missing, pull the
saved stack again while connected, from the install directory (ODS does not
use a top-level docker-compose.yml):
cd ~/ods
docker compose $(cat .compose-flags) pull
# Then disconnect
n8n Workflows Failing
Many n8n integrations require internet (Gmail, Slack, etc.). Use only local-compatible workflows:
- File operations
- Local API calls
- Database operations
- Webhook receivers (internal)
Security Benefits
Air-gapped operation provides:
- Data sovereignty — Nothing leaves your network
- Compliance — Suitable for regulated environments
- Privacy — No usage tracking possible
- Resilience — Works during internet outages
Files Created
ods/
├── .offline-mode # Marker file
├── .env # Updated with offline settings
└── models/
└── embeddings/
└── nomic-embed-text-v1.5.Q4_K_M.gguf
Part of M5: Clonable ODS Setup Server