Skip to content
Ghost in the Droid

🍏 Hello World — On-Device iPhone

Your iPhone talks to itself. In the video, Qwen2.5-1.5B runs inside the Ghost iOS app via llama.cpp on Metal — the agent opens X, reads the feed, and summarizes it without a single byte leaving the handset.

Prerequisites

Section titled “Prerequisites”
  • An iPhone (Apple Silicon iPhones give the best decode speeds)
  • Xcode to build and side-load the Ghost iOS app (see the iOS on-device guide)
  • A .gguf model — Qwen2.5-1.5B is the tested default

Step 1 — Build and install the app

Section titled “Step 1 — Build and install the app”

Follow the iOS on-device guide to build the SwiftUI app and load the model. Two engines are available:

  • llama.cpp on Metal — default, runs any GGUF
  • MLX — opt-in engine for faster decode on Apple Silicon

Step 2 — Airplane mode on

Section titled “Step 2 — Airplane mode on”

Same guarantee as Android: the model, the agent loop, and the screen driving all live in the app. Nothing to phone home to.

Step 3 — Ask, on the iPhone

Section titled “Step 3 — Ask, on the iPhone”

In the Ghost app’s chat:

open X and summarize the posts in my feed

The 1.5B model plans, the agent drives X, reads posts, and streams a summary back into the chat bubble — reasoned on-device, end to end.

What next?

Section titled “What next?”

Try another approach →

Section titled “Try another approach →”

MCP + Claude Code · MCP + Codex · MCP + Antigravity · Ghost CLI · LangChain · Web Chat · On-Device Android · iPhone Duolingo