Ask about what you see
Point at anything on screen, hold the key, ask the question. AIPointer answers about that region without you switching windows or pasting screenshots.
Hold a key. Ask a question. Get an answer about whatever your cursor is pointing at. Free, open source, no telemetry. macOS, Windows, Linux.
No window switching, no pasting screenshots into a chat tab. The answer arrives next to your cursor, then gets out of the way.
Point at anything on screen, hold the key, ask the question. AIPointer answers about that region without you switching windows or pasting screenshots.
Right-Cmd on macOS, Right-Ctrl on Windows and Linux. Works in every app, every window, every time. ESC dismisses it.
Plug in OpenRouter, Anthropic, OpenAI, or Google Gemini. You pick the provider, you pay for your own tokens, you stay in control.
No telemetry, no analytics, no crash reports. Your prompt goes to the AI provider you chose. Nowhere else.
v1.1.7 is the second reliability hotfix on top of v1.1.5: Local TTS now loads cleanly on every Mac (no more "could not load sharp module" on first Read), the Kokoro voice picker remembers your selection after Save, and there's a new ▶ Preview button next to the voice dropdown so you can audition all 28 voices before committing. The 3-engine Voice picker, live model picker, primary + fallback chain, and Finder/Explorer auto-attach from v1.1.5 are all here.
Hit ↻ Refresh models on any provider row and AIPointer pulls the current model list from that provider's API. Pick the one you want — the freshest Gemini, Claude, GPT, or any OpenRouter route — without waiting for an app update.
Choose a primary provider, then arrange up to three fallbacks underneath. Each row can pin its own model. The router walks the chain on transient failures (5xx / 429), so one outage doesn't take you offline. Auth errors still surface so you can fix the broken key.
Read-aloud and voice input work out of the box via your OS voice — no setup, no key, no download. Opt into Local (real Kokoro 82M TTS on your CPU, fully offline + free after a one-time ~231 MB download) or Cloud (OpenAI / Gemini / OpenRouter, max quality) from Settings → Voice. Every engine fails safely into System with a clear inline notice.
Select up to 5 files in Finder / Explorer, press the hotkey — AIPointer reads the selection and attaches them automatically. Or queue manually via the paperclip icon. Images via vision, text / code inlined, PDFs by name. Empty prompt? AIPointer auto-summarizes.
Download the build for your OS. On first launch, paste an OpenRouter, Anthropic, OpenAI, or Google API key into settings. macOS asks for Accessibility and Screen Recording. Both are required.
Hover the cursor over whatever you want help with, then hold Right-Cmd on macOS, Right-Ctrl on Windows or Linux. A glass capsule opens beside your cursor.
Type a question, or use the microphone. Slash commands like /summary, /translate, /code, /identify are available, plain prose works too. Press ESC to dismiss.
The full source is on GitHub under BSL-1.1 (Business Source License). Free for personal, educational, and internal business use. Commercial redistribution, SaaS hosting, and resale require a license. The code is small and deliberate enough to read in an afternoon. Reverts automatically to Apache 2.0 on 2030-05-19.
Today, AIPointer answers what is near your cursor. v1.1.7 is the follow-up reliability hotfix to v1.1.6 — Local TTS now loads cleanly across every Mac variant, the voice picker keeps your saved choice, and a Preview button lets you audition the 28 Kokoro voices before settling on one. The bigger v1.1.5 feature set (3-engine Voice picker, live model picker, primary + fallback chain, Finder/Explorer auto-attach with multi-file upload) is all in. Next: Local STT runtime (whisper.cpp) in v1.2.0, then longer agentic tasks.
Need something bigger? For multi-step agents, computer use, and persistent workflows, point at Skales. Same author, much larger scope.
AIPointer sends nothing to us. Your prompts and screenshots go directly from your machine to the AI provider you chose. No telemetry, no analytics, no tracking, no auto-collected crash reports.
Your API keys are stored in your operating system's secure key store (macOS Keychain, DPAPI on Windows, libsecret on Linux), not in plain files. Sessions stay in memory unless you opt into saving them to disk.
The code is open source, so you can verify all of this for yourself.
Read the full privacy policyAuto-update is built in, so future versions arrive without revisiting this page. SHA-512 hashes are listed for those who verify before installing.
The most common install questions across macOS, Windows, and Linux. More detail and a full feature reference are in the documentation.
Yes — zero clicks. Select up to 5 files in macOS Finder or Windows Explorer, then press the AIPointer hotkey (Right-Cmd / Right-Ctrl) anywhere on your system. AIPointer reads the active file-manager selection at trigger time and auto-queues the highlighted files. The dark-orange “N files attached” pill confirms above the prompt; the cursor screenshot is auto-suppressed so the model focuses on the files.
Hit Enter without typing anything and AIPointer auto-routes
through the /summary template against the files
— useful for “just summarise these” without
having to type anything.
On macOS this works via an AppleScript that guards on Finder
being the frontmost app, so the hotkey from inside Chrome /
VS Code never silently scrapes background Finder selections.
On Windows it queries the foreground Explorer window via
Shell.Application. Linux has no universal
selection mechanism and falls back to the paperclip button.
Supported types: images (PNG / JPG / WebP / GIF — sent as vision context), text / JSON / CSV / Markdown / YAML / XML / code (inlined with file labels), and binary docs (PDF / DOCX / XLSX — referenced by name and size). 20 MB per-file cap.
The Windows installer is currently unsigned, which causes SmartScreen to flag any new app it has not seen reach reputation threshold. The executable is not malicious — Microsoft just has no installation history for it yet.
Click More info, then Run anyway. After running once, SmartScreen quiets down on subsequent updates.
For extra confidence, verify the SHA-256 hash against the value published with the download above before running. The v1.1.7 binary has been submitted to Microsoft Defender Security Intelligence for reputation building; full code-signing will land in v1.2.0.
Notarised builds avoid this warning, but if you grabbed an unsigned build (e.g. from a GitHub Release artefact), right-click the app and choose Open — macOS shows a confirmation instead of refusing entirely.
On first launch macOS asks for Accessibility and Screen Recording permissions. Both are required — Accessibility is how AIPointer detects the trigger hotkey, Screen Recording is how it captures the area around your cursor. Without them the app silently fails to summon.
Make the AppImage executable: chmod +x AIPointer.AppImage,
then double-click or run from a terminal.
AIPointer encrypts your API keys via libsecret (gnome-keyring) or
kwallet. If neither is installed, keys fall back to base64
storage with a warning in Settings. On Debian-based systems:
sudo apt install libsecret-1-0.
Your API key is encrypted by macOS Keychain so it cannot be read out of the config file. The first time we encrypt or decrypt a key, macOS asks for permission. Pick Always Allow so AIPointer can read its own keys silently on subsequent launches.
Since v1.1.0 this prompt is deferred until the exact moment you click Complete in the onboarding wizard — the user-active context where it makes sense.
Child Mode is a behavior layer designed for kids. The Pillbox itself looks identical, but underneath:
Switching back to Adult mode requires a 4–6 digit parent PIN set during onboarding.
Each download row above shows a SHA-256 hash. On macOS run
shasum -a 256 AIPointer-1.1.7.dmg, on Linux
sha256sum AIPointer-1.1.7.AppImage, on Windows
Get-FileHash AIPointer-Setup-1.1.7.exe -Algorithm SHA256.
The output must match the hash on this page exactly.
No. The only outbound network calls are: (a) your queries to the
LLM provider you configured, (b) the auto-updater checking
aipointer.app/updates/ for a new version (toggleable).
No analytics, no crash reporting, no profile syncing.
See the full privacy policy for the long version.