A native macOS AI assistant that reads your current screen and browser context when you ask.

⬇ Download for macOS · Releases · Features · Privacy · Build from source · 繁體中文

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.

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.

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"

• 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.

• 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.

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 and the notices sheet apply the hiding preference when their content attaches to a window. This reduces a timing gap, not the OS-level limits below.
• Enabled by default, this setting requests window hiding through NSWindow.sharingType = .none, while leaving Wisp usable on your own screen. Apple treats this as a legacy mechanism and explicitly says not to rely on it to prevent capture. Results depend on macOS and the recorder's capture path; the switch being on is not proof that the receiving side cannot see Wisp.
• Toggle Try to hide during screen sharing in the menu bar, or use Settings → Permissions → Screen sharing. Turn it off to record Wisp demos.
• Local tests on macOS 26.6.1 hid synthetic windows using Wisp's production hiding code from ScreenCaptureKit display screenshots and video frames. Those same windows remained in ScreenCaptureKit's window list. This is not a compatibility claim for Chrome, Zoom, Meet, Teams, OBS, or a proctoring platform. See the test scope, results, and reproduction steps.
• Menu bar icons and system-owned dialogs are not guaranteed hidden. This setting does not conceal application identity, focus changes, clipboard events, or third-party activity records. Cameras and hardware capture are unaffected. Verify the actual receiving-side view before relying on it.
• Privacy cursor lock (experimental). Settings → Permissions → Screen sharing uses a single private moving arrow after hiding the system pointer. Wisp no longer creates a stationary shared arrow: it could show through translucent windows or remain visible when they moved, producing two cursors. While enabled, the option also uses arrow artwork over text/resize controls and removes custom and plain-button press effects. Turning it off restores native cursor requests and button feedback. Protected panels, settings title bars, non-key floating panels and in-process menus/popovers use the same policy; native drags keep their owner until release. Background hover keeps the system pointer. Clicking activates Wisp before replacing it, preserving the external capture target. Local text selection, insertion carets and menu selection remain usable and visible; their containing windows request exclusion from compatible recordings. The lock defaults off. Enabling it turns on window hiding; turning off window hiding disables the lock, while disabling only the lock keeps windows hidden. Recording tools can still add click circles or transmit pointer metadata independently. Disable click highlighting in the recording tool; Wisp cannot switch it off in other apps. Compatible recordings omit the pointer during Wisp interaction; they no longer receive an artificial parked arrow. See single-cursor verification.
• Resize guide. A thin pale-blue rounded line and small edge markers inside the panel show where the custom resize gesture can start. They are visual hints only; the panel remains clickable; resize cursors become arrows while privacy cursor lock is enabled.

Shortcuts

• The default ⌃⌥Space shortcut remains available and can be changed in Settings → Permissions. Enhanced mode can record Shift, Globe/Fn, modifier-only shortcuts, and double- or triple-tap sequences. Enhanced mode needs Accessibility permission to observe keys while another app is active.
• Available in English and Simplified Chinese. It follows the system language, and Settings → General can pin either language on its own.

Model providers

• Cloud endpoint: sends chat/completions requests with SSE streaming and image_url data URLs, supporting OpenAI, OpenRouter, and other compatible endpoints.
• Ollama: defaults to http://localhost:11434/v1, reads the model list from Ollama, and marks models that appear to support vision.
• Agent CLI: the settings section contains Codex, Antigravity and Claude Code. Codex runs codex app-server --listen stdio:// with an ephemeral, read-only thread; Antigravity and Claude Code use headless stream-json. All three display answer text as it arrives, with screenshots passed through a private temporary workspace. None of them writes session files into Wisp's conversation directory. See CLI streaming and validation for protocol details. Fast is preferred for supported Codex models and supported Claude Opus selections, with standard-speed fallback. AGY 1.1.27 has no Fast switch. Fast can consume more credits; it does not lower reasoning effort or change your selected model. Cloud requests also prefer Fast/Priority on OpenRouter paid models and the official OpenAI and Gemini endpoints. OpenRouter free models keep their free IDs and prefer throughput; unsupported gateways receive no extra parameters. Availability and billing depend on the service. See speed preferences.
• Quick / Deep responses: select a mode above the input (Quick is the default); set each mode's model and shortcut in Settings → Model. Choices remain editable during an answer and apply to the next question. Quick skips new full-page text capture and sends the default model; Deep uses more supported reasoning, and on Antigravity switches to the default model's High thinking level. Both prefer available Fast tiers. A Quick answer can be redone in Deep with one click, replacing it without using another turn. The stopwatch separates preparation, first-text and total latency; higher effort alone does not establish accuracy.
• Commands: saved questions such as Summarize This Page or Translate Selection, each with its own mode and optional shortcut, run from the ✦ menu, the empty chat, or any app. Text in the input box goes with the command as its material. Edit them in Settings → Commands.

Startup and updates

• Optional launch at login through SMAppService, with a direct link to Login Items & Extensions when macOS holds the item for approval.
• Optional update check that asks GitHub once per launch for the latest release tag. It is off by default, sends nothing about you or your usage, never installs anything on its own, and can be switched off entirely.

Local conversations

• Keeps up to 10 conversations by default, with up to 30 user turns per conversation; both limits can be changed in Settings.
• Preserves the latest two turns in full and folds older context into summary rows.
• Limits page text to 60,000 characters, retaining the first 75% and last 25% when the page is longer.
• Stores conversations as readable JSON, decoded tolerantly so that a damaged record or a field added by a future version costs you that record rather than the whole history.
• Refuses to overwrite a conversation file written by a newer version of Wisp, and tells you when a file could not be read instead of quietly starting empty.
• At the conversation limit, offers to drop the least recently updated conversation by name rather than simply refusing to create a new one.
• Settings can delete all conversations and the saved API key.

⌃⌥Space
  ↓ remember the current frontmost app
  ↓ capture the window; read browser URL, title, and page text when available
  ↓ show the panel and label the captured context
  ↓ ask a question → send it to the selected provider
  ↓ keep the conversation locally after the answer completes

Wisp refreshes context at three points: when the panel is shown, when the frontmost app changes while the panel is open, and before sending if the previous capture is older than 20 seconds or you have left the panel. It does not automatically recapture while a response is generating.

Open the menu-bar icon → Settings → Model:

• Cloud endpoint: pick a provider — OpenRouter, Google Gemini, OpenAI, Anthropic, Zhipu GLM, or Custom for any other OpenAI-compatible address — then pick a model and enter that provider's API key. Choosing a provider fills in the Base URL and its model list for you. Keys are saved on blur, Return, provider switch, or closing Settings, one per provider, in the macOS Keychain rather than the conversation JSON, so several providers stay configured side by side and switching never loses one.
• Ollama: start Ollama and refresh the model list. Only a vision-capable model can interpret a screenshot.
• Agent CLI: choose Codex, Antigravity, or Claude Code inside the same section, select the detected executable and optionally choose a model. Wisp uses the CLI login already on your Mac. Antigravity models are refreshed by running agy models, so an Antigravity update does not require a Wisp update.

All three provider groups have Test Connection. Cloud and Ollama tests send a very small test image to each model Quick and Deep use. Codex and Antigravity tests check --version; Claude Code uses auth status so a missing login gets its own message. None of the local CLI checks spends a model request.

These images were generated by the current build's offline rendering entry points. They contain no real desktop, conversation, or personal files. The header image uses simulated Chrome page metadata; the SwiftUI Menu model selector was cropped because it is not rendered by the offline ImageRenderer.

Context header

Persistent island: idle, hover, and generating states

To regenerate the checked-in images, build with the diagnostics flag — the rendering entry points require both DEBUG and an explicit WISP_DIAGNOSTICS, so that neither a shipped app nor an ordinary Debug build can be driven by another local process:

xcodebuild -project Wisp.xcodeproj -scheme Wisp -configuration Debug \
  SWIFT_ACTIVE_COMPILATION_CONDITIONS="DEBUG WISP_DIAGNOSTICS" build

Build/Debug/Wisp.app/Contents/MacOS/Wisp --render-header docs/screenshots/context-header.png
Build/Debug/Wisp.app/Contents/MacOS/Wisp --render-island docs/screenshots/island-states.png

Add -AppleLanguages '(en)' or -AppleLanguages '(zh-Hans)' to render a specific localization.

Local storage

• Conversation text and page-text snapshots: ~/Library/Application Support/Wisp/conversations.json.
• API keys: stored in the macOS Keychain, outside Wisp's support directory.
• Screenshots: never written to the conversation JSON. Cloud and Ollama keep them in memory; local CLIs may write them to a private per-request temporary directory and remove it when the command ends. A crash can defer cleanup to macOS's temporary-directory maintenance.
• Optional debug files: when debug capture is enabled, Wisp writes debug/last-context.json and debug/last-screenshot.jpg under its application-support directory.

Network boundary

• The OpenAI-compatible provider sends the context you selected and the current screenshot to the Base URL you configured. The service's logs, retention, and privacy policy are outside Wisp's control.
• Ollama uses localhost by default. If you configure a remote Base URL, the request goes to that address.
• Wisp starts a local Codex app-server, creates a temporary directory for image input, uses an ephemeral thread and a read-only sandbox, and removes the temporary directory when the command ends. Codex's own account, network, and service-side logging are outside Wisp's control.
• Wisp starts Antigravity locally, in its own temporary working directory, with --sandbox and Antigravity's JSON headless mode. An attached screenshot is written into that directory and named in the prompt so Antigravity reads it from disk; the directory is removed when the request ends. Antigravity's account, network, quota, and service-side logging remain outside Wisp's control.
• Wisp starts Claude Code in its own temporary working directory with --restricted --tools Read --no-session-persistence. Only the screenshot files placed in that workspace are available to its file tool; Wisp removes the directory when the command ends. Claude Code's account, network, quota, and service-side logging remain outside Wisp's control.
• If the update check is enabled, Wisp asks api.github.com once per launch for the latest release tag. The request carries no identifier beyond your IP address and a Wisp/<version> user agent, downloads nothing, and installs nothing. Turning it off makes no request at all.
• Wisp has no account system, sync service, analytics SDK, crash-reporting SDK, or automatic background recording. Explicit Live listening sessions continuously capture selected audio until stopped; see details.

Exclusions are per app, not per site. The exclusion list takes bundle identifiers, so there is currently no way to exempt one URL or domain while still using Wisp in that browser. Entire-screen captures always cut out excluded apps, plus any apps you check under Settings → Capture → Capture range.

The full policy is in PRIVACY.md.

Permissions

• Screen Recording: current-window or entire-screen screenshots.
• Automation / Apple Events: browser URL and title access, plus page JavaScript execution for supported browsers.
• Accessibility: only for the enhanced shortcut mode, to observe Shift, Globe/Fn, modifier-only, or double-/triple-tap key events while another app is active. The standard shortcut does not need this permission.
• Network client: cloud endpoints, remote Ollama, or networking performed by Codex itself.

The current release is 0.4.0 (build 6); see CHANGELOG.md. It corresponds to Git tag v0.4.0.

The Universal 2 package passed lipo -info checks for both architecture slices. Release validation was performed on Apple Silicon; an Intel hardware runtime regression has not yet been completed. The package is ad-hoc signed and not Apple-notarized, so the first launch may require Control-click → Open.

defaults write com.yichenlin.Wisp AppleLanguages -array en

~/Library/Application Support/Wisp/

xcodegen generate
xcodebuild -project Wisp.xcodeproj -scheme Wisp -configuration Debug build
cp -R Build/Debug/Wisp.app "$HOME/Applications/"

tools/install-local.sh              # build, sign, install; clean temporary backup after verification
tools/install-local.sh --no-install # build and sign only
tools/install-local.sh --keep-backup # optional: retain the rollback copy

xcodebuild -project Wisp.xcodeproj -scheme Wisp -configuration Debug \
  CODE_SIGNING_ALLOWED=NO build

rm -rf Build
xcodebuild -project Wisp.xcodeproj -scheme Wisp -configuration Release \
  -arch arm64 -arch x86_64 \
  ONLY_ACTIVE_ARCH=NO \
  CODE_SIGN_IDENTITY=- CODE_SIGNING_REQUIRED=YES \
  CODE_SIGN_INJECT_BASE_ENTITLEMENTS=NO build

mkdir -p dist
lipo -info Build/Release/Wisp.app/Contents/MacOS/Wisp
ditto --norsrc -c -k --keepParent \
  Build/Release/Wisp.app \
  dist/Wisp-macOS-universal.zip
shasum -a 256 dist/Wisp-macOS-universal.zip > dist/Wisp-macOS-universal.zip.sha256

plutil -p Build/Release/Wisp.app/Contents/Info.plist
codesign --verify --deep --strict Build/Release/Wisp.app
codesign -d --entitlements :- Build/Release/Wisp.app
unzip -l dist/Wisp-macOS-universal.zip
shasum -a 256 -c dist/Wisp-macOS-universal.zip.sha256

• Wisp/App/ — app entry point, menu-bar lifecycle, global shortcut, and diagnostic entry points.
• Wisp/Capture/ — screen capture, browser AppleScript, page text, and context orchestration.
• Wisp/LLM/ — OpenAI-compatible HTTP, Ollama, Agent CLI, SSE parsing, and prompt assembly.
• Wisp/Store/ — local conversation JSON and macOS Keychain access.
• Wisp/UI/ — floating panel, persistent island, chat, context header, conversation list, and Settings.
• Wisp/Support/ — permissions, screen geometry, UserDefaults settings, login item, and update checking.
• Wisp/Resources/ — the string catalog.
• WispTests/ — unit tests for CLI isolation, event and auth parsing, prompt truncation, completion checks, and temporary-directory cleanup.
• Wisp/Assets.xcassets/ — macOS app icon and image assets.
• docs/screenshots/ — the current offline-rendered README screenshots.
• project.yml — XcodeGen project source, version settings, dependencies, localization configuration, and signing configuration.
• LICENSE, THIRD-PARTY-NOTICES.txt, PRIVACY.md, CHANGELOG.md — licence, bundled dependency notices, privacy policy, and release notes. The notice is copied from the repository root into the app bundle at build time.

project.yml is the source of truth for MARKETING_VERSION and CURRENT_PROJECT_VERSION; Wisp.xcodeproj is regenerated with XcodeGen. The current release convention is:

1. Update the marketing version or build number in project.yml.
2. Run xcodegen generate and complete a Debug or Release build.
3. For a distributable build, verify both architecture slices, bundle metadata, signature, ZIP integrity, and checksum.
4. Create a vX.Y.Z Git tag and upload Wisp-macOS-universal.zip plus its .sha256 file to the matching GitHub Release.

Build/, dist/, Xcode user state, diagnostics, and local environment files are ignored by Git. Source, icon assets, project files, the package resolution, and public documentation are tracked.

• The current public package is ad-hoc signed and not Apple-notarized.
• The Universal 2 slices have been generated and checked on Apple Silicon, but the current release has not yet been run on an Intel Mac.
• Browser extraction depends on supported bundle identifiers, Automation permission, browser JavaScript settings, and page security boundaries.
• Codex streaming requires a CLI with the app-server v2 protocol (verified with 0.153.4). Each request still carries Codex's own fixed context cost.
• Antigravity streams agent response deltas (verified with 1.1.27); older versions that emit only a final result display that result at completion. Each request carries Antigravity's own fixed context cost — previously measured at about 30,400 tokens before any of Wisp's content. Screenshots are passed as files because Antigravity's headless input accepts text only — it rejects an image_url content block — so a screenshot request spends one extra tool turn on reading the file, which costs about 1,150 tokens rather than inlining the image. Antigravity also truncates its own input at roughly 71,400 tokens without saying so, so Wisp fits the prompt to a smaller budget first and marks what it left out; a page far larger than the default 60,000-character limit therefore reaches Antigravity abridged, with the gap declared, rather than silently incomplete.
• CI covers unit tests, Universal 2 compilation, signature verification, and the Release entitlement check. There is still no public notarization and release-signing pipeline.
• Because releases are ad-hoc signed, every build has a new code identity. macOS ties Screen Recording, Automation, and Keychain access to that identity, so updating the app can require granting those permissions again.
• App exclusions are per bundle identifier. There is no per-URL or per-domain exclusion, which is the exclusion most useful in a browser.
• Conversation history is stored as unencrypted JSON and is not evicted by age. Conversation and turn counts are bounded, but message bytes are not strictly capped; the file is rewritten in full on every message. Live captions do not append messages automatically.
• The island can be dragged only in its desktop form; the notch form stays anchored to the notch.
• The island's circle can reach a screen edge but not overlap it, so its centre stops one radius (20pt) inside the edge.

Wisp bundles KeyboardShortcuts 2.4.0 by Sindre Sorhus, under the MIT License. Its full notice ships inside the app — Settings → General → Open-source licenses — and is checked in as THIRD-PARTY-NOTICES.txt.

Everything else is written from scratch in SwiftUI and AppKit, with no other third-party dependencies.

Wisp is copyright © 2026 YC and available under the MIT License.

• GitHub repository
• Latest release
• Release assets
• Issues
• Release notes
• Privacy policy
• Third-party notices
• Traditional Chinese README