jacky 0.8.2

A local, privacy-first, English-only voice assistant.

Jack

Your Mac, with a voice. Entirely on-device.

Website · Docs · Download

---

Jack is a privacy-first voice & chat assistant for macOS. You talk to a floating orb (or type in a chat drawer); it listens, understands, and acts on your Mac — open apps and websites, find and open files, check battery/Wi-Fi/disk, remember things about you, empty the Trash. Everything runs locally: no audio, text, or memory leaves your machine, with two clearly-marked, opt-in exceptions (web search and an optional cloud LLM). English only.

Highlights

• Chat or voice — type in the chat drawer (the default), or enable voice and say “jack, …” for hands-free (push-to-talk also supported).
• Acts through a permission gate — read-only tools run straight through, destructive ones ask first, and every action is audited. The model plans; the gate executes — it never runs a tool blindly.
• Hands on your Mac — app & web control, on-device file search (open / reveal in Finder), system info, and local memory.
• Local by default, cloud optional — runs fully on-device with Ollama, or switch to Claude (disclosed, opt-in).

→ Full feature tour, the orb's states, and how it works: jack's website.

Quickstart

Download (recommended): grab the latest .dmg from Releases — a single, self-contained app. It's an unsigned dev preview, so the first launch is right-click Jack → Open. Voice models download on demand the first time you enable voice.

Build from source (macOS, Apple Silicon):

brew install uv ollama
ollama serve && ollama pull qwen3:8b   # local LLM, one time
make setup                              # dev env + git hooks
make run                                # the engine/daemon

Then build the orb from ui/orb-shell (cargo tauri dev). Full setup, configuration, permissions, and troubleshooting live in Getting started.

Install the coding CLI

No Python needed — one command downloads a self-contained jack binary.

macOS / Linux

curl -fsSL https://raw.githubusercontent.com/mdjakkariya/jacky/main/install.sh | sh

Windows (PowerShell)

irm https://raw.githubusercontent.com/mdjakkariya/jacky/main/install.ps1 | iex

Then run jack inside a project. You'll need an LLM backend: set a cloud key with jack config (Anthropic/OpenAI — you pay the provider), or install Ollama and pull a model to run fully local.

Update: jack checks once a day and prints a notice when a new version ships — run jack update to upgrade in place. jack --version shows the current version.

Architecture

A headless Python engine + a macOS orb / chat client (Tauri) that talks to it over a localhost WebSocket. A pipeline of swappable stages drives one turn:

AudioSource ─▶ SpeechToText ─▶ LanguageModel ─plan─▶ PermissionGate ─▶ ToolRegistry
 wake+VAD/PTT   whisper         Ollama / Claude       risk·confirm·audit  apps·web·files·system
                                                                                  │
                              TextToSpeech (Piper) ◀──────── reply ◀──────────────┘

Diagram: docs/architecture/architecture.svg. Engineering conventions and the full design rationale are in CLAUDE.md.

Star history

Contributing

Contributions are welcome — see CONTRIBUTING.md for the workflow (open an issue first, make check, Conventional Commits, DCO sign-off) and CODE_OF_CONDUCT.md.

Sponsor

Jack is built in my free time. If it's useful to you, sponsoring helps me keep going via GitHub Sponsors — or buy me a chai:

License

Created and maintained by Mohamed Jakkariya (@mdjakkariya). Licensed under Apache-2.0 — see LICENSE and NOTICE. You're free to use, modify, and build on Jack; please keep the attribution.