Pointr needs an AI model to work at all, and offers more that make it more capable. All are configured from the tray icon → Settings.
1
Connect a Gemini API key (free) or an OpenAI API key, whichever you'd rather use. One of the two is required - you can also connect both and switch anytime, and pick any model your key supports, not just the default.
2
Optionally connect Tavily for live web search, and GitHub if you want Pointr reading your repos, issues, and pull requests.
3
Press Ctrl + Alt + Space anywhere to start asking. The first time, Pointr asks once whether to share anonymous usage data. Skipping is fine and changes nothing else.
Requires 64-bit Windows 10 (version 1903 or later) or Windows 11, plus the Microsoft Visual C++ Redistributable, which most PCs already have. If Pointr won't start and mentions a missing MSVCP140.dll or VCRUNTIME140.dll, install the x64 redistributable from Microsoft and try again.
Choose Gemini, OpenAI, or both
Gemini API key
Pointr uses Google's Gemini models to read your screen and answer questions - this is one of two supported providers, not the only option. Every request runs on your own free API key, never a shared one. Pointr's backend relays each request to Google and keeps nothing afterwards (see Privacy). Prefer OpenAI, or want to use a different model than the default? See OpenAI and model choice.
How to get one
1.
Go to aistudio.google.com and sign in with a Google account.
2.
Click "Get API key", then "Create API key". The free tier is enough for normal use.
3.
Copy the key - it starts with AIza…
4.
In Pointr, open Settings → Gemini API key, paste it in, and click Connect.
Choose Gemini, OpenAI, or both
OpenAI and model choice
Gemini and OpenAI are both fully supported, everywhere Pointr uses a model: answers, explain mode, agent tasks, multi-step tasks, and the GitHub, web search and file tools. Connect either key, or both and switch between them anytime, under Settings → AI model. Pointr defaults to gemini-3.1-flash-lite for Gemini and gpt-6-luna for OpenAI, but the model dropdown isn't limited to those - once a key is connected it lists every model that key can actually use, and a "Custom model ID" box takes any model name at all.
To use OpenAI
1.
Create a key at platform.openai.com/api-keys. OpenAI has no free tier for the API, so usage is billed to your own OpenAI account.
2.
In Pointr, open Settings → OpenAI API key, paste it in, and click Connect.
3.
Under AI model, switch the provider to OpenAI and pick a model - this is also where you switch back to Gemini later, or change either one's model.
The same steps work for picking a different Gemini model: no OpenAI key needed, just open AI model with Gemini already selected and choose from its own model list. Note: clicking on screen is tuned for Gemini's coordinate format, so if an OpenAI model keeps clicking in the wrong place, switching back to Gemini is the fix.
Optional
Tavily API key
Tavily gives Pointr live web search, so agent tasks can pull in current information instead of relying only on what's on your screen. Skip this if you don't need Pointr searching the web.
How to get one
1.
Go to tavily.com and create a free account.
2.
Copy the API key from your dashboard - it starts with tvly-…
3.
In Pointr, open Settings → Tavily API key, paste it in, and click Connect.
Optional
GitHub connection
Lets Pointr read repositories, issues, pull requests, and commits on your behalf when you ask it about GitHub. Read-only - Pointr never pushes, comments, merges, or opens anything on GitHub for you. Only needed if you plan to ask Pointr about your code.
How to get one
1.
Go to github.com/settings/personal-access-tokens/new to create a fine-grained token, not a classic one.
2.
Under Repository permissions, grant read-only access to Contents, Issues, and Pull requests. Don't grant any write permissions - Pointr doesn't need or use them.
3.
Copy the token - it starts with github_pat_…
4.
In Pointr, open Settings → GitHub, paste it in, and click Connect.
Audio & voice
Read answers aloud
When on, Pointr speaks each answer as it arrives, in addition to showing it. Turn this off if you'd rather read silently.
Narration voice
Picks which system voice reads answers and storyboard narration aloud. Only applies while "Read answers aloud" is on - use the Test button in Settings to preview a voice before saving.
On-screen actions & permissions
Settings → On-screen actions decides whether Pointr may act on your computer, and whether it asks first. Click any tier below to explore how Pointr handles permissions:
Recommended🛡️
Always ask
Shows plan or keystroke proposal in an overlay. Waits for explicit Enter before touching any window.
Default • Full Safety
Autonomous⚡
Always allow
Starts straight away upon plan generation. Each step still displays live on screen. Esc halts at any time.
Fast Multi-Step Execution
Read-Only🔒
Off
Completely restricts Pointr to answering questions and drawing screen annotations. Never types or clicks.
Strict Zero-Action Mode
Permission & Execution Lifecycle
Phase 1
Analyze Screen
Reads focused app & form fields
➔
Phase 2: Gate
User Verification
Enter executes • Esc cancels
➔
Phase 3
Live Execution
Automated typing & clicks with audit log
Single-action vs Multi-step tasks
For single steps (e.g. typing a draft into Slack), Pointr waits for Enter. For complex multi-step tasks, Pointr presents its entire breakdown plan upfront. Once confirmed with Enter, it carries out each step smoothly while observing the screen before each click.
Agent history audit log
Every multi-step run is logged so you can see what Pointr did and whether each step worked. Open it from the tray icon → Agent History, or the clock button beside the question box. It keeps your last 50 runs, only on your own computer, and holds no screenshots. Use Clear history to wipe it.
Browser tasks
Needs Node.js
For tasks like "open a browser and search for X", Pointr opens its own browser window and works the page by the names of its buttons, links, and fields instead of guessing where to click. That makes web tasks much more reliable than clicking by position.
1
Install Node.js (the standard installer is fine). This is the only thing you set up yourself.
2
Run your first browser task. Pointr downloads its browser component the first time, about 300 MB. It needs an internet connection, takes a minute or two, and only happens once.
It's a separate browser
The window Pointr opens is a fresh one, not your everyday browser. It isn't signed in to anything and doesn't have your bookmarks or extensions. It stays open when a task finishes so you can see the result, and closes if you cancel the task.
Known limits
Links that open in a new tab aren't followed yet, and plain drop-down menus may not respond. If a task gets stuck on one of these, Pointr stops and says so rather than looping.
Voice Flows & Privacy Architecture
100% On-Device Available
Hold Ctrl + Win, say what you want, and let go. Pointr captures your screen the moment you start, shows your words as you speak, and sends them the instant you release, exactly as if you'd typed them. Choose between local on-device transcription and cloud-assisted transcription:
Input
Push-to-Talk
Hold Ctrl + Win
Mic active only while held
➔
100% On-Device
NVIDIA Parakeet
Local 670 MB Speech Model
Audio never leaves your PC
➔
Memory Clean
Zero Disk Audio
Audio discarded instantly
Only text prompt forwarded
1
Open Settings → Voice and click Download to fetch the on-device model (670 MB once). Or switch provider to My OpenAI key to skip the download entirely.
2
Hold Ctrl + Win anywhere on your screen. When Windows asks, allow microphone access for desktop apps. Release the keys when done speaking.
Your voice stays on your computer
Speech is turned into text on your own machine by NVIDIA's Parakeet model. Audio is never uploaded or saved; only the resulting text is sent, the same as a typed question. The microphone is on only while you hold the keys.
Good to know
It understands English and 24 other European languages. It uses about 1 GB of memory while voice is on; turn voice off in Settings to free it. Pressing another key while holding Ctrl + Win (for example Ctrl + Win + D) cancels listening and does the normal Windows shortcut instead. Holding the keys while Pointr is reading an answer aloud stops it and starts a new question.
Anonymous usage data
Off by default. Pointr asks once, the first time you use it, and you can change your answer any time in Settings under "Share anonymous usage data".
What's sent, if you turn it on
Which features get used, how many steps a task took, how it ended (finished, stopped, or errored), the Pointr version, your Windows build number, and a random ID generated on your computer. Pointr also tells the analytics service not to record your IP address or location.
What's never sent
Screenshots, your questions or task text, clipboard contents, file names or contents, Pointr's answers, and API keys or tokens. The full picture of what leaves your computer, and where it goes, is on the Privacy page.
Keywords & agentic flows
Pointr reads plain questions by default. A couple of keywords at the start of your prompt switch it into a different mode - narrated walkthroughs or multi-step agent tasks.
explain: <topic>
Storyboard walkthrough
Breaks the answer into narrated steps, speaking each one aloud while drawing point, box, and arrow annotations directly on your screen. Good for "explain this codebase" or "explain how this chart works."
agent: <task>
Multi-step agent task
Hands the request to Pointr's agent loop: it can read your clipboard and screen, search the live web when Tavily is connected, read your GitHub repos/issues/PRs when connected, and propose typing into a field or opening an app. A single action stops for your confirmation first, and multi-step tasks show a plan you confirm once (see On-screen actions). Nothing is ever written back to GitHub. Try "agent: summarize the open issues in this repo" or "agent: what's the latest stable Node version."
no keyword
Direct ask
Anything typed without a keyword is answered directly from what's visible on your screen - the default, fastest path for quick questions.
drag a region
Region select
Not typed - held via Ctrl + Alt + Shift + Space. Draw a box around any part of the screen to scope a question to exactly that chart, error, or paragraph.
Keyboard shortcuts
Pointr has three ways to ask a question, plus a couple of keys used while an answer or action is on screen.