✈ Airplane Mode

One-stop setup for offline local LLM coding on your Mac. Built for flights and no-WiFi situations — get a real coding agent running locally with one command.

Bundles the official Ollama runtime, Qwen3 models, and the OpenCode harness, with environment settings tuned for Apple Silicon.

Note: The default model choices and tuned settings in this repo were dialed in for a 14" MacBook Pro (M-series, 36GB unified memory). They're a solid starting point for similar Apple Silicon Macs, but if you have more or less memory, adjust the model size and context accordingly — see GUIDE.md for the memory math.

---

⚠️ Security first — read this

A coding agent runs with your user permissions. It can read/write files and execute shell commands. Treat every component — the harness, the model, and especially any plugins / extensions / skills — as a potential attack surface and supply-chain risk.

Airplane Mode deliberately bundles only:

• Ollama — official installer (not Homebrew)
• Qwen3 — models from the official Ollama library
• OpenCode — open-source harness by SST

It installs no third-party plugins, extensions, MCP servers, or skills. If you add any yourself, you own that risk. For work on sensitive or proprietary repositories, clear agent usage with your security/AppSec team first, and consider running the agent inside a container/sandbox.

---

Quickstart

# 1. 🛜 Clone + run once on good WiFi (acknowledges a security notice, then sets up)
git clone https://github.com/daniilrrr/airplane-mode.git
cd airplane-mode
bash ./src/airplane-mode.sh                    # default: Ollama engine
# or, for the power-user path:
bash ./src/airplane-mode.sh --engine llama.cpp

# 2. ✈️ Any time you want to code offline:
airplane-mode                 # launches OpenCode + your model
airplane-mode qwen3.6:27b     # (ollama engine) pick a different model

Engines: Ollama vs. llama.cpp

Pick your runtime with --engine:

Engine	--engine	Best for	Trade-off
Ollama (default)	ollama	Easiest setup, ollama pull, auto-wired ollama launch	An abstraction layer; less control
llama.cpp	llama.cpp	Max control, raw tuned flags, fewer black-box surprises	More setup; you manage GGUF + flags

Note: Ollama is itself a wrapper around llama.cpp — the actual open-source inference engine that does the work. The --engine llama.cpp path talks to that engine directly (full credit upstream), runs a tuned llama-server, and points OpenCode at its OpenAI-compatible endpoint. Same models, fewer abstractions.

Before you fly: turn WiFi OFF on the ground and run a real task. If it works dry, it works at 40,000 feet.

---

What the script does

Step	Action	WiFi?
1	Security acknowledgment gate (type yes)	—
2	Verify macOS + Apple Silicon	—
3	Install official Ollama (rejects the broken Homebrew build)	🛜
4	Start the Ollama server in offline mode	—
5	Pick your model (defaults to qwen3-coder:30b) and pull it	🛜
6	Write tuned env vars + an airplane-mode launcher to your shell profile	—
7	Install OpenCode	🛜
8	Smoke test (loads the model on GPU, up to a 3-min timeout)	—

Everything after the pulls works fully offline.

---

Model-by-task

Model	Use for	Size
qwen3-coder:30b	Coding — repos, edits, PRs, refactors (default)	19GB
qwen3.6:27b	Reasoning / chat / planning	17GB

Use the coding model for repo work (it caches cleanly and skips slow "thinking" blocks); use the reasoning model for non-coding tasks. Switch with airplane-mode <model>.

---

Requirements

• macOS Sonoma (14) or newer
• Apple Silicon (M-series) strongly recommended — Metal GPU acceleration
• ~20–40GB free disk (per model)
• Enough unified memory: the model + its KV cache must fit. 16GB is tight; 24GB+ comfortable; 36GB+ ideal. Remember your IDE and other apps eat into this — see GUIDE.md for the memory math.

Reference machine for the defaults: 14" MacBook Pro, M-series, 36GB unified memory.

---

Tuned environment variables

The script writes these to your shell profile:

Variable	Value	Why
OLLAMA_CONTEXT_LENGTH	64000	Coding/agents need a big context window
OLLAMA_FLASH_ATTENTION	1	Cuts memory growth as context grows
OLLAMA_KV_CACHE_TYPE	q8_0	~half the KV cache memory, ~no quality loss
OLLAMA_KEEP_ALIVE	-1	Keep the model pinned in memory
OLLAMA_NO_CLOUD	1	Force true offline — never phone home

---

Full guide

See GUIDE.md for the complete walkthrough: install details, the memory budgeting math, performance tuning, troubleshooting (including the "llama-server not found" Homebrew gotcha), harness comparison, and in-flight survival tips.

---

Troubleshooting (quick hits)

• llama-server binary not found / 500 errors → you have the Homebrew Ollama. brew uninstall ollama and re-run the script. (Don't run the cmake command the error suggests — that's a red herring.)
• Slow responses (1–2 min) → usually the harness's huge system prompt + a model's "thinking" mode, not raw speed. OpenCode + qwen3-coder:30b is the snappy combo. See GUIDE.md → Performance.
• Running on CPU instead of GPU → the model doesn't fit in memory. Close apps, use a smaller model, or lower the context. Confirm with ollama ps → PROCESSOR = 100% GPU.

---

License

MIT