⬡ Ghostbar

The only native Swift macOS Cluely alternative.

No Electron. No subscription. No cloud. Just Swift.
AI invisible to screen sharing — Zoom, Teams, Meet, OBS.

Download DMG · Build from source · Report bug

---

Why Ghostbar over Pluely, Natively, or Vysper?

Every other Cluely alternative is built on Electron or Tauri — that means a bundled Chromium browser, 200–500MB of RAM just to open, and slow startup times.

Ghostbar is written in native Swift using AppKit. It starts instantly, uses minimal memory, and feels like a real macOS app — because it is one.

	Ghostbar	Pluely	Natively	Vysper
Tech	✅ Native Swift	Tauri/Rust	Electron	Electron
macOS only	✅ Purpose-built	❌ Cross-platform	❌ Cross-platform	❌ Cross-platform
Ollama / local LLMs	✅	⚠️ Partial	✅	❌
Free	✅ Always	✅	✅ Free tier	✅
No subscription	✅	✅	⚠️ Premium tier	✅
Open source	✅ MIT	✅	✅	✅
No telemetry	✅ Zero	⚠️	⚠️ Limited	❓
Voice input	✅ On-device	❌	✅	❌
App size	✅ ~5MB	~10MB	~150MB	~200MB

---

What is Ghostbar?

A macOS menu bar AI client that is completely invisible to screen capture.

Zoom, Google Meet, Microsoft Teams, OBS, QuickTime, Cmd+Shift+5 — none of them see it. It only exists on your physical display.

// The entire secret. One native AppKit API.
window.sharingType = .none

No hacks. No injection. A public, documented Apple API that removes the window from the display capture pipeline before any recording tool can touch it.

---

Use cases

Technical interviews
LeetCode, HackerRank, take-home assessments. Ask for hints, complexity analysis, edge cases — all while sharing your screen. The interviewer sees your code. They don't see Ghostbar.

Live coding demos
Presenting to a client or team? Use AI to look up syntax, generate boilerplate, or sanity-check logic in real time. Nobody notices.

Work calls & meetings
Prepare answers on the fly. Summarize what was just said. Draft a response before you speak.

System design interviews
Ask for architecture patterns, trade-offs, scalability approaches instantly.

Local & private
Pair with Ollama or LM Studio for fully on-device inference. Nothing leaves your machine.

---

Tested invisible in ✅

Tool	Status
Zoom	✅ Invisible
Google Meet (Chrome)	✅ Invisible
Microsoft Teams	✅ Invisible
OBS Studio	✅ Invisible
QuickTime screen recording	✅ Invisible
macOS Cmd+Shift+5	✅ Invisible

---

Features

🫥 Invisible by default	Hidden from every screen capture tool on macOS
🍎 Menu bar only	No Dock icon. No trace. Appears as ⬡
🤖 Multi-backend	Ollama, OpenAI, Claude, OpenRouter, NVIDIA NIM, LM Studio, llama.cpp
🎙 Voice input	On-device transcription via whisper-cpp
📸 Screenshot analysis	Capture your screen → attach to message. Model sees it, recorder doesn't.
🔒 Zero telemetry	No analytics, no cloud, no tracking
⚡ Instant startup	Native Swift — no Chromium, no Electron overhead

---

Install

Option A — DMG (fastest)

Download Ghostbar-v1.0.1.dmg, drag to Applications, open.

Option B — Build from source

# Requirements: macOS 13+, Xcode Command Line Tools
xcode-select --install

git clone https://github.com/rbc33/Ghostbar.git && cd Ghostbar
bash build.sh
open Ghostbar.app

The ⬡ icon appears in your menu bar. Click → Open chat.

---

How the invisibility works

macOS exposes a window-level API that controls whether a window participates in the display server's capture pipeline:

NSWindow.sharingType = .none     // excluded from all capture
NSWindow.sharingType = .readOnly // default — visible to recorders

Setting .none tells the macOS compositor to exclude the window from all capture operations before any recording application, screenshot tool, or API (CGWindowListCreateImage, SCStreamConfiguration, etc.) can observe it.

The window renders normally on your physical display. It simply does not exist to capture pipelines.

---

Backends

Open ⬡ → Settings… and pick your backend:

Backend	URL	Notes
Ollama	http://localhost:11434	Local, no key
OpenAI	https://api.openai.com/v1	API key required
Anthropic	https://api.anthropic.com/v1	API key required
OpenRouter	https://openrouter.ai/api/v1	API key required
NVIDIA NIM	https://integrate.api.nvidia.com/v1	Free tier available
LM Studio	http://localhost:1234/v1	Local, no key
llama.cpp	http://localhost:8080/v1	Local, no key

Any OpenAI-compatible server works — select OpenAI and set the URL.

Free models via NVIDIA NIM

1. Go to build.nvidia.com → sign in → any model → Get API Key
2. Key starts with nvapi-
3. In Ghostbar: select NVIDIA NIM, paste key

Hundreds of models (Llama, Mistral, Gemma, Qwen, DeepSeek) — no credit card required.

---

Voice input

brew install whisper-cpp
curl -L -o /opt/homebrew/share/whisper-cpp/ggml-base.bin \
  https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.bin

Press 🎙 or ⌘⌥ → speak → press again → transcribed and sent. Fully on-device.

Model	Size	Quality
tiny	75 MB	Fast
base	150 MB	Balanced ✓
small	470 MB	Better
medium	1.5 GB	Best

---

Screenshot analysis

⌘⇧ → captures screen → attaches to next message.

The model sees your screen. The screen recorder doesn't see Ghostbar.

Works with any vision model: llava, gpt-4o, gemma3, claude-opus-4-7…

---

Keyboard shortcuts

Action	Shortcut
Send message	Enter
New line	Shift+Enter
Voice input	⌘⌥
Screenshot	⌘⇧
Close	Cmd+W

All shortcuts customizable in Settings.

---

Privacy

• No telemetry. No analytics. No crash reporting to any server.
• No cloud. Network calls go only to your configured backend.
• Local backends (Ollama, llama.cpp, LM Studio) run entirely on your machine — nothing leaves it.
• Voice transcription runs on-device via Whisper.
• Screenshots are sent only to your configured backend.

---

Contributing

PRs welcome. Open an issue first for large changes.

git clone https://github.com/rbc33/Ghostbar.git
cd Ghostbar
open Package.swift   # Xcode opens automatically

---

Made with ⬡ — because some things should stay invisible.