Why

• On the device. MLX, llama.cpp or Apple Foundation Models run the model. Prompts, tools and answers stay on the phone.
• The phone is the toolbox. Image and voice generation, transcription, translation, vision, torch, haptics, deep links — one line hands them to the model.
• A small, honest loop. Model → tools → model, until the model answers. Errors go back to the model; approvals pause the loop. Swift 6, Foundation only at the core.

Quickstart

.package(url: "https://github.com/cagataycali/loopl", from: "0.1.0")

import Loopl
import LooplRuntimes
import LooplAppleTools

let model = MLXModel()                                   // or GGUFModel(), AppleFoundationModel()
try await model.load(spec, from: weights) { _ in }

let agent = Agent(model: model, tools: AppleTools.available)   // what this iPhone can do
for try await event in agent.stream("Torch on, and how do I say 'where is the exit?' in Turkish?") {
    if case .text(let chunk) = event { print(chunk, terminator: "") }
}

Get started · Write a tool · API reference

The app

The SDK, shipped as a free iOS app. Everything below runs on the phone.

• Models — MLX and GGUF from Hugging Face; add a token to list and download your private repos.
• Share — select conversations, push them to a private Hugging Face dataset.
• Train your own — launch a Hugging Face job from the phone, see the cost first, get the model back in Models.
• External drive — keep models on a USB-C drive when the phone is full.
• Voice and vision — record with the mic and the model calls transcribe; attach a photo for vision models.
• Tools — torch, haptics, open apps, image and voice generation, translation; risky calls ask first.

Models

Up to 4B at 4-bit fits an 8 GB iPhone. Any public MLX or GGUF repo loads the same way. All models →

The loopl models — Qwen3.5 0.8B/2B/4B post-tuned for this loop, vision kept, with their cards — are public, in one Hugging Face collection.

How it works

Model → tools → model, until the model answers. No iteration cap. A failed call returns to the model as an error result, so it can fix the call; a tool that needs approval pauses the loop until you answer. The agent loop →

Build from source

swift test                                                     # SDK, macOS
brew install xcodegen && xcodegen generate -s App/project.yml  # the app → App/LooplApp.xcodeproj
scripts/fetch-llama-xcframework.sh --with-sim                  # once, for GGUF

Docs

loopl.dev · Get started · The app · Models · FAQ · Changelog

_{Made by @cagataycali.}