MLXKit

Swift helpers for building local LLM apps on Apple platforms using the mlx-swift ecosystem.

MLXKit focuses on the annoying parts you hit immediately when building a real app:

• Download MLX Community models from Hugging Face
• Store + list downloaded models (in your app’s Documents/models/…)
• Load an MLXLLM model container
• Chat + stream tokens and optionally handle tool calls

This package is used by my example agent app ComfyPilot (browser-controlled agent + in-app model download + tool calling).

Requirements

• Swift tools: swift-tools-version: 6.2
• Platforms (as currently set in Package.swift):
  • macOS 26+
  • iOS 26+

Install (Swift Package Manager)

Add MLXKit as a dependency in your app’s Package.swift:

.package(url: "<your-repo-url-for-MLXKit>", branch: "main")

Then add the product to your target:

.product(name: "MLXKit", package: "MLXKit")

MLXKit itself depends on:

• mlx-swift
• mlx-swift-lm

Quick Start

1) Download + Select A Model (UI-Friendly)

Use ModelLoaderService to:

• list downloaded models
• download a new model by name (from mlx-community)
• prompt the user before downloading the full file set

import MLXKit

@MainActor
let loader = ModelLoaderService(selectFirst: true)

// Refresh local models from Documents/models
loader.sync()

// Download a model from mlx-community/<name>
Task {
  await loader.download(named: "Llama-3.2-3B-Instruct")
}

On macOS you can also open the models folder:

#if os(macOS)
loader.openModelFolder()
#endif

Where models are stored:

Documents/models/<model-name>/...

In Your App Container, For Example: ~/Library/Containers/com.whatever.app/Data/Documents/models

2) Load The Model

import MLXKit

@MainActor
let chat = MLXChatService()

if let model = loader.selected {
  await chat.loadModel(
    at: model.url,
    defaultPrompt: "You are a helpful assistant."
  )
}

3) Stream A Response (+ Tool Calls)

MLXKit streams assistant tokens via completion and surfaces tool calls via toolcallCompletionHandler.

import MLXKit
import MLXLMCommon

let tools: [[String: any Sendable]] = [
  [
    "type": "function",
    "function": [
      "name": "search",
      "description": "Search the web for information",
      "parameters": [
        "type": "object",
        "properties": [
          "query": ["type": "string"]
        ]
      ]
    ]
  ]
]

let messages: [ModelMessage] = [
  .init(role: .user, content: "Search for the weather in Chicago.")
]

let _ = try await chat.getResponse(
  messages: messages,
  tools: tools,
  completion: { tokenChunk in
    // Append streamed tokens to your UI
    print(tokenChunk, terminator: "")
  },
  toolcallCompletionHandler: { toolCall in
    // Route tool calls to your app (browser, network, filesystem, etc.)
    print("Tool:", toolCall.functionName, "args:", toolCall.arguments)
  }
)

What’s Included

• MLXChatService
  • loads an MLX LLM container
  • streams responses
  • emits tool calls
• ModelLoaderService
  • lists installed models
  • downloads models from mlx-community on Hugging Face
  • supports a “confirm before download” UI flow
• MLXChatModel
  • lightweight model folder reference (Documents/models/<name>)

Notes / Caveats

• Model storage is currently based on URL.documentsDirectory (iOS/macOS sandbox documents).
• The Package.swift platform minimums are currently set to v26. If you need older OS support, you’ll want to lower those and verify Observation/@Observable usage. I do it cuz I have had no need to support older Versions.
• Please Open a issue if you need an older OS Version.