LocalPilot

Offline AI coding assistant for VS Code — private, air-gapped, self-hosted

How to run a fully offline AI coding assistant in VS Code with LocalPilot — no internet, no accounts, no telemetry. Works on air-gapped machines and in regulated environments.

Ask AI about this page
ChatGPT Claude Perplexity Gemini View as Markdown

Teams in finance, healthcare, defence and anyone with an NDA need an assistant whose data path they can audit. LocalPilot’s data path is one line long: VS Code → your model server.

Zero-network checklist

  1. Model server on the same machine or LAN — Ollama, LM Studio, llama.cpp llama-server or vLLM. All run without internet once models are downloaded.
  2. LocalPilot installed from the VSIX — no activation call-home, no update checks, no telemetry. Verify with SHA256SUMS.txt from the release.
  3. Web tools off — set localpilot.approvals.network to never to hard-block web_search, web_fetch and MCP network tools.
  4. Approvals on — keep localpilot.approvals.terminal at ask so every shell command is reviewed.
  5. Audit what changed — the Agents panel lists every file each session touched; Undo restores pre-agent content.
// settings.json for an air-gapped workstation
{
  "localpilot.providers": [{ "id": "ollama", "type": "ollama", "baseUrl": "http://127.0.0.1:11434" }],
  "localpilot.approvals.network": "never",
  "localpilot.approvals.terminal": "ask",
  "localpilot.webSearch.provider": "searxng"   // irrelevant when network is 'never', kept explicit
}

Sizing hardware

LocalPilot does this for you: Settings → Models reads RAM, cores and GPU memory, shows a model budget, and marks every catalog size fits / tight / too big. Automatic setup starts with the smallest useful model and leaves bigger ones as an explicit upgrade. As a rule of thumb:

Setup Comfortable models
8 GB laptop qwen2.5-coder:3b + qwen2.5-coder:1.5b for completions
MacBook Pro M-series 16–24 GB qwen2.5-coder:7b, qwen2.5-coder:14b (q4), gemma3:4b for screenshots
32–64 GB unified / 24 GB GPU qwen2.5-coder:32b, devstral, qwen3:32b, gpt-oss:20b
Shared GPU server (vLLM) 32B–72B models for the whole team, OpenAI-compatible endpoint

Pre-download models before going offline (ollama pull … or the catalog’s Install button); the Ollama installer itself needs network once.

Team roll-out

Next: Configuration · Parallel agents & change review