AICompleteChat

A full-source, production-quality example of a fully on-device AI chat app for macOS. No cloud dependency once the model is downloaded — chat, voice dictation, and memory all run locally.

Just want to try it? Skip building from source — grab the signed, notarized build from the latest release, unzip, and run it directly. No DesignFoundationPro license needed for this — that's only required to build from source (see below).

What it does

• On-device inference via MLX (mlx-community/gemma-4-e4b-it-4bit), downloaded from Hugging Face the first time you launch it and cached locally after that — every launch after the first is fully offline.
• Persistent chat history — conversations survive across launches, with delete support.
• Voice dictation and commands via AiVoiceKit — on-device speech recognition, global hotkeys, AI-assisted rewriting of selected text in any app.
• Persistent memory via AiPersona — a bi-temporal knowledge graph that extracts facts and entities from conversations, retrieves relevant ones per turn, and lets you browse, correct, or delete anything it's learned (Settings → Memory → Browse memory). It also remembers who you are: set your name once and it becomes a real fact in the graph, not just a settings field.
• A UI assembled from a real design system, not built from scratch — see below.

Requirements

• macOS 15+, Apple Silicon (MLX doesn't run meaningfully on Intel)

• Xcode 26+

• xcodegen (brew install xcodegen) — this repo has no committed .xcodeproj; regenerate it from project.yml

• A DesignFoundationPro license. This is the one intentional gate: the source here is meant to be read and learned from by anyone, but it only actually builds for someone with access to DesignFoundationPro, since the entire UI is assembled from its AIChat vertical rather than built from scratch.

  Without access, package resolution fails outright — you'll see something like this during xcodegen generate + build, or from Xcode's own package resolution:

  error: Failed to clone repository https://github.com/NerdSnipe-Inc/DesignFoundationPro.git:
      remote: Repository not found.
      fatal: repository 'https://github.com/NerdSnipe-Inc/DesignFoundationPro.git/' not found
  

  (If you have no GitHub credentials configured in git at all, the underlying git error reads could not read Username for 'https://github.com': terminal prompts disabled instead — same cause, just failing one step earlier.) That's expected, not a bug in this repo: it means your GitHub account isn't on DesignFoundationPro's access list. Grab a license to fix it, or just download the signed build above instead.

Every other dependency is public and resolves automatically via Swift Package Manager: DesignFoundation (the free design-system core), AIChatKit / AIChatKitMLX (chat session + MLX provider), AiPersona (memory graph), and AiVoiceKit (voice).

Building

git clone https://github.com/NerdSnipe-Inc/AICompleteChat.git
cd AICompleteChat
xcodegen generate
open AICompleteChat.xcodeproj

Build and run the AICompleteChat scheme. First launch downloads the model from Hugging Face (several GB) — you'll see real download progress, not just a spinner.

Architecture

AICompleteChat/
  ContentView.swift          — bridges AIChatKit's ChatSession to DesignFoundationPro's AIChat UI
  Engine/
    AppEnvironment.swift     — owns the MLX provider, chat session, memory store, voice engine
    PersonaChatCoordinator.swift — folds persona identity + retrieved memory into every prompt
  Core/
    ChatHistoryStore.swift   — SwiftData-backed chat persistence
  Settings/                  — voice model picker, persona editor, memory browser

The UI itself lives in DesignFoundationPro's AIChat vertical (DFAIChatRootView and friends) — this app wires real data into it rather than reimplementing chat UI from scratch. See DesignFoundationPro's own docs for how that vertical is put together.

Known limits

• First launch needs the model. The app uses mlx-community/gemma-4-e4b-it-4bit (about 5 GB; the local snapshot is 4.8 GB), downloaded on first launch by AIChatKitMLX's MLXModelManager, which retries with backoff, waits while offline, and resumes an unfinished download after a quit (AIChatKitMLX 1.4.0+). The FunctionGemma router model, used only when tool routing is enabled, is not covered by that and downloads on first use. The live tests are skipped, not failed, when it isn't cached. See docs/TESTING.md.
• FunctionGemma tool routing is experimental and off by default. Enable it with defaults write com.nerdsnipe.aicompletechat AICOMPLETECHAT_TOOL_ROUTING -bool YES. Read AIChatKitMLX's TOOL_ROUTING.md first for the measured accuracy and latency.
• Router benchmarks are opt-in. The router accuracy/latency live tests only run with ROUTING_BENCHMARK=1; the other routing tests run whenever both models are cached.
• A message that got no reply is not sent to the model on later turns. This covers a cancel before any output and an error or empty reply before any output. It stays in the transcript, shown dimmed with a “Cancelled” or “Not answered” caption (UserEntry.isCancelled / UserEntry.isFailed; the error/empty-reply case is AIChatKit 2.0.0+). Resend it if you still want an answer.

Building from source also requires DesignFoundationPro access, see Requirements.

License

MIT — see LICENSE. Note this only covers the code in this repo; DesignFoundationPro, AiPersona, and AiVoiceKit are each licensed separately (see their own repos).