Air-Gapped Local LLM Setup (Ollama / vLLM)

For defense, healthcare, and enterprise banking environments where source code and inference traffic cannot leave the corporate firewall, ScanDrix supports 100% offline, air-gapped operation using local model servers like Ollama or vLLM.

Recommended hardware & models

1. Hosting with Ollama

Run Ollama on an internal server accessible from your ScanDrix worker instances:

# Pull and start high-performance coding model
ollama pull qwen2.5-coder:32b
OLLAMA_HOST=0.0.0.0:11434 ollama serve

In .scandrix.yaml or your ScanDrix environment configuration:

ai:
  provider: custom_openai
  endpoint: "http://ollama-service.internal:11434/v1"
  model: "qwen2.5-coder:32b"
  api_key: "ollama"  # Ollama requires a non-empty placeholder string

2. Production hosting with vLLM

vLLM provides paged-attention and continuous batching, serving dozens of concurrent pull request reviews with minimal latency:

python3 -m vllm.entrypoints.openai.api_server \
  --model Qwen/Qwen2.5-Coder-32B-Instruct \
  --tensor-parallel-size 2 \
  --port 8000 \
  --max-model-len 8192 \
  --gpu-memory-utilization 0.95

Point ScanDrix to your internal vLLM cluster:

ai:
  provider: custom_openai
  endpoint: "http://vllm-cluster.internal:8000/v1"
  model: "Qwen/Qwen2.5-Coder-32B-Instruct"
  api_key: "env:INTERNAL_VLLM_AUTH_TOKEN"
  max_tokens: 4096
  temperature: 0.1

Verifying air-gapped network isolation

To verify that ScanDrix makes zero outbound connections to external AI APIs:

1. Disable public internet egress on your worker VPC subnet (0.0.0.0/0 -> blackhole).
2. Run a test review via the CLI:
   bashCopy

   scandrix review --staged
   

3. Check the ScanDrix worker audit logs to confirm that all completion tokens were served from your internal inference IP with 0 bytes transmitted to external hosts.