Privacy Policy

Last updated: 3 November 2026

Navisual (“we”, “us”) is a Windows desktop app that watches your active window and shows you where to click. This policy explains what data the app handles, what leaves your computer, and the third parties involved. The short version: your screenshots are processed to answer your request and are not stored on our servers, your saved sessions stay on your machine, we log the task you type (you can switch that off, and nothing is logged when the AI runs on your own machine), and the free tier needs no sign-up.

Navisual is operated from British Columbia, Canada. Because the AI and infrastructure providers below operate globally, data you send (such as a screenshot for a request) may be processed on servers outside Canada, including in the United States.

1. What the app processes

• Screenshots of the window you point it at. When you ask for guidance, a screenshot of the application window you are being guided through is sent to the AI provider you’ve selected so it can identify the next step. You can choose to share a whole screen instead, from the app picker inside Navisual; if you do, that is what is captured and sent, and it stays that way until you choose a single window again. If you turn on Autopilot, Navisual also checks the screen on your computer about twice a second while it is guiding you, and when the screen has changed enough to look like you finished the step, it sends a new screenshot for the next step without you asking. The check itself happens on your computer and sends nothing; the screenshot it leads to is sent like any other. The AI may also ask to see a screenshot from an earlier step of the same task again, for example to read a number an earlier page showed; if it does, that screenshot is sent again with the current request. To make that possible, Navisual may keep the screenshots of a task on your computer until the task ends; nothing is captured for this that was not already sent. Any screenshot Navisual saves is saved on your own computer; it goes nowhere except back to the AI as just described, or into a file you export yourself, and is never stored on Navisual’s servers.
• Window titles, and your Windows version. The focused window’s title is sent with the request to give the AI context. So are the titles of the windows from the earlier steps of the same task, each with what you said at that step, so the AI can follow a task that moves between windows or pages. Also sent: which version of Windows you are running and its build number, for example “Windows 11 (build 26200)”. Windows versions differ enough in Settings and File Explorer that without it the guidance has to guess which one you are looking at. When a step asks you to press a keyboard shortcut, Navisual checks on your computer whether another program has already claimed that shortcut, and if so tells the AI only that it is taken, not which program took it. No other details about your computer are sent: not your hardware, not your drives or free space, not your installed software. (Earlier versions also sent the window’s on-screen position and size; this one does not.)
• What the window’s controls say about themselves. Navisual reads the names, types and states of the buttons, menus, tabs, lists and fields in the window you are being guided in, from Windows’ accessibility interface (the one screen readers use): for example Button “Bold”, on, or a drop-down showing “14”. That list is sent with the request so the AI can name exactly the control to use. It never includes what you type into a text box.
• Your request text. What you type or say, plus a short running summary of the task, and short notes the AI chose to keep from an earlier screen of the same task (a code or a number that screen showed, for example). Those notes are sent again with the later requests of that task, and are saved with the session on your machine.
• Your saved sessions (local only). The conversation, your task, the plan, what you clicked and each step’s screenshot are saved on your machine so you can reopen recent sessions; the screenshots can be switched off in Settings. The most recent are kept and older ones go to the Recycle Bin. You can export a session to a file in any folder you choose (the file includes the screenshots, if they were kept), and open such a file again later. Opening one shows it; it rejoins the recent list only if you carry on working in it. These files never reach Navisual’s servers (the task you type is logged on its own, see §4).
• Learned notes about each app (kept on your machine). When finding a command in an app took more than one try, Navisual writes down the names of the menus and buttons you clicked to get there, with one line the AI writes about it, in a small note for that app. The note is kept on your computer and sent with your later requests in the same app, so the guidance gets it right the first time. It holds interface labels only, never file names, document text, addresses or long numbers. You can read, change or delete these notes in Settings ▸ Screen Guide ▸ Learned notes. They are never stored on Navisual’s servers (on the managed tier they pass through our relay with the request, like the screenshot).
• The control you last clicked. While a guidance session is active, Navisual records the name and type of the control you click (for example Button “Save”), so the AI knows what you just did instead of having to guess it from pixels. This applies only to clicks inside the app you are being guided in; clicks anywhere else on your screen are discarded immediately and are never stored or sent. The click is also kept with your saved session on your machine, so a reopened session can show what you clicked. Navisual records a control’s name, never its contents: the text inside a box is not read, and a password field is reported only as “(password field)”. Navisual does not monitor your keyboard at all.
• Where your cursor is, in supported apps. Navisual reads the cursor’s page, section and line number, whether text is selected, and the style and text of the paragraph you are in (long ones are shortened). Those words are already in the screenshot you send. A screenshot cannot show a text cursor, so without this the AI cannot tell where you are in a document.
• Voice input (optional). If you enable push-to-talk, your spoken audio is transcribed by the WebView2 Web Speech API, which streams it to Microsoft’s online speech service. This is a separate data flow you can leave disabled.

Cursor position stops the moment guidance stops and is never written to disk unless you turn on prompt logging in Developer settings. Click context is kept with your saved sessions (the most recent ones, on your machine). Neither is collected while capture is paused with the Pause hotkey.

Element coordinates are matched locally on your machine (Windows UI Automation + built-in OCR). Your screen layout is not used to build any profile of you.

2. Where your screenshot goes (you choose)

The AI provider is your choice in Settings:

• Managed tier (free / paid): routed through our relay (see §5) to an AI provider on your behalf.
• Bring your own key: sent directly to Anthropic, Google, OpenAI, DeepSeek, or Qwen using your own API key; our servers are not involved in the request (we do log the step-outcome rows of §4).
• Ollama (local), or Custom (OpenAI-compatible) pointed at a local server such as LM Studio or llama.cpp: the model runs on your own machine or LAN, and your screenshots and requests never leave your network, and we log nothing about your use either (§4).

Whichever provider receives a request processes it under their privacy policy, and we don’t control how third-party AI providers use it.

Important: the free managed tier uses your data for model training. The free tier is served by free-of-charge AI models: we route to whichever provider(s) currently offer the best reliability and cost for the free tier, and that routing changes from time to time as we improve it. Free-of-charge AI access is commonly offered on the condition that the provider may retain your requests, including the screenshots of your active window, and use them to train and improve their AI models. This is standard practice across the industry for no-cost API tiers, not something specific to whichever provider we currently route to. If you do not want your screen content used for model training, use one of the private options instead:

• the paid managed tier: routed to paid provider endpoints, whose current API policies state they do not use your data for model training (this is their policy, not a guarantee by us, and a provider may change it, see their terms);
• a bring-your-own-key provider on a paid plan, under that provider’s current policy (note: a provider’s own free tier (e.g. a free Google AI Studio key) may itself train on your data; check that provider’s policy); or
• Ollama, or Custom pointed at a local server (local): your screenshots and requests never leave your machine, so no provider policy applies to them at all.

Some providers apply different data-use rules by region, for example certain providers extend paid-tier-equivalent data protections to free-tier users in the European Economic Area, Switzerland, or the UK, under that provider’s own policy. Check the current upstream provider’s own terms for the specifics that apply to your region.

3. Optional app add-ons (Blender)

For a few applications whose interface Navisual cannot read reliably from the screen alone, we offer an optional add-on that lets the app tell Navisual where its buttons are. Today this exists for Blender. It is entirely optional: Navisual works without it, and nothing about it runs unless you install and enable it yourself.

• How it gets installed. Navisual can copy the add-on file into Blender’s add-ons folder for you, but only when you click Install on the offer. It is then inert until you tick its checkbox in Blender’s own Preferences → Add-ons panel: that checkbox is the consent step, and Navisual cannot tick it for you.
• What it does when enabled. It listens on a local network port (127.0.0.1:47611) that is reachable only from your own computer: it does not accept connections from other machines and does not send anything to the internet. Navisual asks it read-only questions such as “where is the tool shelf?”, “which mode is open?”, and “which brush is selected?”.
• What it can’t do. It is read-only by construction: it cannot click, type, run commands, or change or save your file. This matches Navisual’s core rule: the AI guides, it never acts.
• What reaches the AI. The interface facts above (mode, tool, brush names, object names in your scene) are included in the request to your chosen AI provider, alongside the screenshot, and are covered by §2. No file contents, geometry, or file paths are sent.
• Turning it off. Untick it in Blender’s Add-ons panel to stop it, or use Remove there to delete the file. Navisual keeps working without it.

4. Data we store

On your computer, in %LOCALAPPDATA%\com.navisual.app: your settings, conversation/session files (including each step’s screenshot, unless you turn that off), the learned notes about each app, a diagnostic log of what the app did (it can include the AI’s summary of your task and the names of the apps you were guided in), and, only if you turn on debug logging, more detailed diagnostic files. API keys you enter and your sign-in token are not in that folder: they are kept in Windows Credential Manager, encrypted by Windows for your user account. (Only if Windows refuses to store one is it written to the settings file instead, so you are not signed out.) None of this is uploaded, except what a request needs (an API key goes to its own provider, the sign-in token to our servers, the learned notes to the AI, see §1), the task you type on the step-outcome rows below, or what you send us yourself.

On our servers (Supabase). Each item says which providers it applies to.

• Account record (made the first time you use the managed tier, which is the default): an anonymous user id, your request count, coin balance, tier, and (after a purchase) your Stripe customer id. The free tier uses anonymous sign-in, so it holds no name or email until you register (Google or email) to buy coins.
• Free-tier device count (managed tier only): to keep the 30 free requests per device (rather than per throwaway anonymous session), the app sends a one-way hash of a Windows machine identifier with managed-tier requests, and we store that hash alongside a request count only to enforce the free limit. The hash cannot be reversed to identify your machine or you; paid requests bill by coins and ignore it, and bring-your-own-key or local providers never receive it.
  
  Alongside it we record the IP address that device was first seen from. It is written once, when a device is new to us, and never updated afterwards: it answers “where did this device first appear”, not “where is it now”. Its only use is anti-abuse: it lets us notice one network minting many fresh free allowances, and refuse them. It is not used to locate you, and it is kept only on that device’s row: never on your account record, a relay record or a step-outcome row, and never with what is on your screen. The relay records below do carry the same device hash for 90 days, so for that long the two could be matched.
• Relay records (managed tier only): our relay, the server that passes a managed request on to the AI, logs one short row each time it does: which provider and model answered, whether a fallback was used, how long it took, the outcome (served, provider error, quota reached, refused), and your anonymous user id and device hash. It is how we notice a provider failing or requests being refused for the wrong reason. This row holds no screenshot, none of your words, and no instruction: the detail it keeps is status codes. (The task you typed is on the step-outcome row below, unless you switch that off.) Deleted after 90 days.
• Step outcomes and feedback (any provider, except an AI on your own machine or network, see below). We log one short row each time you ask for guidance and each time a step ends, so we can see how often the guidance is right and what people use Navisual for. No screenshots, ever.
  
  When a row is logged. Each time you ask for guidance (when the answer comes back, or the request fails); when you move to the next step (→ Next, or an Autopilot advance); when the AI says the task is done; and when you report a step with ✗ This is wrong.
  
  What every row holds: the outcome (asked, moved on, done, or the reason you picked), the name of the app being guided, the on-screen target the step named, whether the pointer was placed and how it was found, the provider and model, the app version, a session id, and your account's anonymous id if the app has one (see Account record above). Only on a ✗ This is wrong report: the instruction the AI gave. What you typed, unless you switch it off (below): the task, on any row, and on a ✗ This is wrong report any note you typed with it.
  
  What you type is your own words, so read this bit. It is logged as you wrote it, trimmed to 500 characters, which means anything you put in the box is logged with it: a file name, a link, a customer's name. Don't type something into Navisual you would not want us to hold. You can switch it off in Settings ▸ Screen Guide ▸ What you type; the rest of the row is still logged.
  
  If the AI runs on your own machine or network, we log nothing at all: no rows, whatever that setting says. Navisual looks at the address it is calling: a model on localhost, or on a machine at a private address such as 192.168.x.x, answers you without anything leaving your network, and Navisual keeps it that way. With your own API key for a provider such as Anthropic or OpenAI, your request already goes over the internet to them, so rows are logged as described above.
  
  How long we keep it: your words (the task and any note) for 12 months, the rest of the row for 24 (§8).

5. Third-party processors

6. International data transfers

Our managed services run on infrastructure provided by Supabase and the upstream AI providers listed above. Your account record, the other records of §4, and the request payloads you send (screenshots of your active window plus your text) may be transferred to and processed in the United States or other jurisdictions where these providers operate. By using the managed tier, you consent to this transfer. If you prefer that your screenshots and your words never leave your machine, use a local provider: Ollama, or Custom (OpenAI-compatible) pointed at a local server. Your screenshots and what you type then go nowhere: not to us, not to any AI provider, and we log nothing about your use (§4). Leave voice input off as well: it goes to Microsoft whichever provider you use (§1).

7. Payments

Coin purchases are processed by Stripe. We receive a confirmation that a payment succeeded (to credit your balance) and store your Stripe customer id; we do not receive or store your card number. Stripe’s handling of your payment data is governed by Stripe’s Privacy Policy.

8. Retention & deletion

On Navisual’s servers, the relay passes managed-tier requests (the screenshot and the text) on in transit and does not keep them after the response; what we do keep is listed in §4. Our hosting provider (Supabase) also keeps short technical logs of the connections to our servers (times, addresses, status codes and your IP address), for troubleshooting and to stop abuse, and deletes them automatically after a few days. Upstream AI providers, however, may retain them: the free tier’s models may keep your request and use it to train their AI (see §2), while the paid tier’s providers, under their current paid-API policies, state they do not use it for training (their policy, which they may change). Relay records are deleted after 90 days. Anything you wrote on a feedback row (the task you typed and any note you added) is erased after 12 months, and the row itself is deleted after 24. Your own words have the shorter life of the two, because what is useful to us about an old row is the outcome, not the wording. Your account record persists until you ask us to delete it: deleting your account removes the account record, and detaches your id from any feedback or relay records that remain, so what is left is no longer linked to you. Anything you typed on a feedback row stays on it until its 12 months are up, account or not. The device row behind the free limit (§4) is keyed to a machine rather than an account, so it is not reached by account deletion; ask us (§11) if you want it cleared too. Local files live on your machine until you uninstall or delete them. To request deletion of your managed-tier account data, contact us (§11).

9. Your rights

Depending on where you live (e.g. GDPR in the EU/UK, CCPA in California), you may have the right to access, correct, or delete the data we hold about you, and to object to or restrict its processing. The free tier needs no name or email, so we hold neither until you register; but the IP address a device was first seen from (§4), and anything you put in a task you type, can identify you. Contact us to exercise any of these rights.

10. Children

Navisual is not directed to children under 13 (or the minimum age in your country), and we do not knowingly collect their data.

11. Contact

Questions or requests: support@navisualguide.com, or open an issue at github.com/NavisualGuide/navisual.

12. Changes

We’ll update this page when our practices change and revise the “last updated” date above.

See also our Terms of Service.