Wisp

Wisp lives in the menu bar. Press ⌃⌥Space, and it remembers the frontmost app before its panel appears. It can then capture the current window (or, if you choose, the whole screen) and, when the frontmost app is a supported browser, read the URL, title, selected text, and page body. Ask a question without copying context between apps.

It is a local-first desktop shell around the model provider you choose: OpenAI-compatible HTTP, Ollama, or Agent CLI (Codex, Antigravity, and Claude Code). Conversation text and page-text snapshots are stored on your Mac; local CLIs may place the current screenshot in a private, per-request temporary directory that Wisp removes when the command ends. The interface ships in English and Simplified Chinese and follows your system language.

Current distribution status: the latest release is v0.4.0 (build 6), a Universal 2 build with arm64 and x86_64 slices. Releases are ad-hoc signed and not Apple-notarized, so the first launch may require Control-click → Open.

What gets sent: when you use a cloud provider, the full text of the page you are looking at and a screenshot of the current window (or the whole screen, if you choose Entire screen) go to the endpoint you configured. Exclusions are per application, not per site — see PRIVACY.md.

Live listening

Voice input shares the main chat panel: a single live caption line with the elapsed time, the microphone/stop button, Stop & analyze, and the rest under a ⋯ menu. Configure source/application/language in Settings → Audio. Control–Option–R starts/stops recording; stopping stages recent speech. Control–Option–D stages it while capture continues. Edit, then Enter sends through the selected AI with normal screen context. For a meeting that just ended, Stop & analyze (Control–Option–Return) stops, waits for the trailing text, and sends it in one key; Control–Option–A does the same without stopping. One more button swaps the answer area between the AI answer and this session's raw transcript. Nothing is submitted without one of those explicit actions, and no completion notification is emitted. Settings → Data can clear listening records separately. The archive retains its 2 GiB / 100-session admission budget. Hiding the panel stops recording. Settings → Audio also carries a master switch: turn voice input off and the row leaves the panel, the shortcuts stop starting a recording, and a session still running is stopped.

Settings → Audio → Recognition engine lists Apple Speech plus every local sherpa-onnx model found under ~/Documents/huggingface/models/ (SenseVoice and Qwen3-ASR architectures today). Copy another model folder there and click Rescan; no reinstall is needed. Local output is written in Simplified Chinese by default (Qwen3-ASR sometimes answers in Traditional); Settings → Audio → Chinese script changes it. See engine setup and limits.

Hold to ask (Settings → Audio → Shortcuts, unassigned by default): hold the shortcut, say your question, release. It is sent with the current screen context in the current response mode. Microphone only, never saved to listening records, and unavailable while a recording runs. A tap shorter than 0.3 s sends nothing; after 60 s the words wait in the input box instead of being sent.

Device/language support is required for on-device transcription. Audio selection is application-level, not per participant or browser tab. This feature does not promise undetectability. See setup, storage limits, and verification scope.

If the menu-bar icon is missing, reopen Wisp from Finder or Spotlight to reveal the panel; its gear always opens Settings. Settings → Privacy controls intentional hiding. System removal no longer changes that preference: Wisp retries insertion once, then shows the panel if removal repeats. Full-screen menu hiding, limited menu-bar space and third-party menu utilities remain controlled by macOS/the utility.

Quick start

1. Download Wisp-macOS-universal.zip and unzip it.
2. Move Wisp.app to ~/Applications or /Applications.
3. On first launch, Control-click Wisp.app, choose Open, and confirm. The public build is not notarized, so a regular double-click may be blocked by Gatekeeper.
4. Grant Screen Recording permission in System Settings. The first time Wisp reads a browser page, grant Wisp Automation access to that browser.
5. Open Settings → Model, choose a provider (changes save automatically), and click Test Connection.
6. Return to the window you want to ask about and press ⌃⌥Space.

If Control-click → Open is unavailable, clear the quarantine flag:

xattr -dr com.apple.quarantine "$HOME/Applications/Wisp.app"
open "$HOME/Applications/Wisp.app"

System requirements

• macOS 14.0 or later.
• An Apple Silicon or Intel Mac. The published app is a Universal 2 binary.
• Screen Recording permission for current-window or entire-screen screenshots.
• Automation permission and the browser's Allow JavaScript from Apple Events setting for full-page browser text.
• Network access and your own API key for cloud endpoints.
• A running Ollama service for the Ollama provider.
• A locally installed and authenticated Codex CLI, Antigravity CLI or Claude Code for the Agent CLI provider.

Why Wisp

• Context is captured before the panel opens. Wisp records the target app first, so the assistant panel does not accidentally become the subject of its own screenshot.
• The request boundary is visible. The header shows the current app, browser information, screenshot and page-text status, and conversation count.
• Capture is on demand. The always-available island tracks the current app but does not continuously record the screen or run browser scripts.
• The model connection is yours. Use a cloud-compatible endpoint, a local Ollama model, or your existing Agent CLI login through Codex, Antigravity, or Claude Code.
• Conversation retention is explicit. Conversation and turn limits are configurable, and deletion is initiated by the user rather than hidden automatic cleanup.

Features

Screen and browser context

• Capture the current frontmost application's window; screenshots are normally kept in memory only.
• Or capture the entire display that window is on, to ask about several windows at once. Excluded apps, and any running apps you check, are cut out of it.
• Supports Chrome, Brave, Edge, Vivaldi, Yandex Browser, Opera, Safari, Arc, and selected stable or beta bundle identifiers.
• Read the current URL, page title, selected text, and page body from supported browsers.
• Report cross-origin iframe URLs and unavailable body text so the model does not assume that a page was fully read.
• Fall back explicitly to URL and screenshot context when JavaScript is disabled or the current app is not supported.
• Virtual-page extraction clears its temporary window.__wispCollector state after the read; Wisp reads Chromium preferences to check the Apple-Events setting but does not write Wisp data into the browser profile.

Floating panel and island

• A menu-bar app with no regular Dock window.
• Starts as a compact card and expands upward when needed; the conversation area scrolls within a bounded height.
• Supports dragging, resizing, Escape to collapse, and automatic collapse after leaving the panel.
• Optional launch at login, so the island is there after a restart.
• The persistent island shows the current app and context status. In its desktop form you can drag it anywhere on screen — it collapses to its small circle while you drag, so edges stay reachable — and it stays there across launches. Expanding follows the space available, growing inward from either edge instead of always from the middle. It can also snap to the Mac's camera notch.

Screen-sharing visibility (best effort)

• The macOS privacy research and expanded capture matrix covers independent-process capture, live filter changes, native video recording, and alternatives such as protected video layers and a separate filtered share window. Information popovers