AI That Can See Your Screen: 6 Tools Compared

The short answer

Pick the capture model before you pick the brand.

If you want help with whatever is open right now, use an on-demand assistant. HeyClicky is the clearest Mac option for spoken guidance and agent tasks. GhostLy is the smaller Mac overlay for people who want to connect their own Claude, GPT or Gemini account. Crade covers Mac and Windows and can move from answering to clicking and typing.

If you want to find something you saw yesterday, that is a different product. Screenpipe continuously builds a local work history across Mac, Windows and Linux. Windows Recall does a narrower version on supported Copilot+ PCs.

AYO is the interesting third route. It says it reads screen text with the same accessibility layer used by screen readers, without saving or uploading screenshots. It was still Windows-only early access when we checked, so treat it as a watchlist option rather than something you can deploy to a team today.

How we checked

AC0 already ranked for ten versions of this question in Search Console, including "AI that watches your screen", "on screen AI helper" and "AI that reads screen". Together they produced 36 impressions and no clicks over the 12 months ending August 16, 2026. Most average positions sat between 26 and 51, and there was no page answering the whole decision.

We checked each product's own site, documentation, repository and privacy policy on August 16, 2026. We did not install and test all six, so feature and privacy statements below are the vendor's documented behaviour, not a lab result. Prices and early-access status move. Follow the source link before buying.

AC0 sells consulting, not screen assistants, and has no paid relationship with any company in this comparison.

The three ways an AI can see your screen

The phrase "watches your screen" hides the most important difference.

Capture model	What happens	Best for	Main risk
On demand	The app reads the current screen after a hotkey or question	Help with the app in front of you	A visible secret can be included in the request
Local screen text	Accessibility APIs extract buttons, labels and text without keeping a screenshot	Voice help with less image exposure	The resulting text or summary may still go to a cloud model
Continuous memory	The app records screen changes over time and builds a searchable history	Finding past work, calls and decisions	A much larger archive exists on the device

Local capture and local inference are not the same thing. An app can read the screen locally, then send a summary to a cloud model for the answer. Another can keep its database on your laptop but use Claude or ChatGPT when you search it. The useful privacy question is not "is it local?" It is:

1. What does it capture?
2. When does capture happen?
3. What is stored?
4. What leaves the device?
5. Which company receives it?

Six tools compared

Tool	Best for	OS	Capture and storage	What can leave the device	Can it act?	Status and price checked August 16
HeyClicky	Learning an unfamiliar Mac app	macOS 14.2+	On-demand screenshot. Vendor says screenshots are not stored; basic text summaries are kept	Screenshots, transcripts and prompts go through HeyClicky's proxy to third-party AI providers	Yes, agent tasks	Free tier; Pro $20/mo; Max $100/mo
GhostLy	Bring-your-own-model help on Mac	macOS	Screen text processed locally. No prompt storage by GhostLy	Prompt goes to your selected provider; visual questions may include a screen capture	Answers, not desktop control	20 free queries; $20 one-time licence
Crade	Cross-platform help plus actions	macOS, Windows	Active-screen snapshot per request; screenshot not retained by Crade after the reply. Text conversations are saved	Snapshot and prompt go to Crade's model or directly to your connected provider	Yes on Agent mode	Free; Pro $7.99/mo; Premium $19.99/mo
AYO	Windows voice assistant with local screen reading	Windows 10/11	Vendor says no screenshots, audio or raw screen text are saved. Short summaries stay local	The question and short context summary go to AWS Bedrock for reasoning	Vendor says yes	Early-access waitlist; final paid pricing not published
Screenpipe	Searchable memory and approved workflows	macOS, Windows, Linux	Continuous screen, app text and audio history in a local database by default	Nothing with local models; optional cloud models and sync can receive selected context	Automations and integrations, not general UI control	Free entry option; paid plans. Commercial source use requires a licence
Windows Recall	Finding something seen earlier	Windows 11 on a Copilot+ PC	Opt-in snapshots saved and indexed locally, encrypted and protected by Windows Hello	Snapshot history is not shared with Microsoft by default; online actions can send selected content	Limited Click to Do actions	Built into supported PCs; still marked preview

HeyClicky: best for being shown where to click on a Mac

HeyClicky is built for the moment when you are staring at an unfamiliar panel in Figma, DaVinci Resolve or another desktop app. Press a hotkey, ask out loud, and it sees the current screen. It can answer, draw on the interface and point at the place it means. Its separate agent mode can take a longer task away and work on it.

The capture boundary is unusually easy to understand. HeyClicky's FAQ says it only sees the screen when you press the hotkey and does not store screenshots. It does keep basic text summaries for context. Its privacy policy adds the network detail: screenshots, transcripts and prompts pass through HeyClicky's backend proxy to third-party providers including Anthropic, OpenAI, Deepgram and Cerebras. On-demand is not the same as on-device. Close password managers, customer records and private messages before asking.

The free tier is enough to learn whether the interaction feels natural. The current paid tiers are $20 per month for unlimited talk and dictation with 150 agent messages, or $100 per month with 1,000 agent messages. It is Mac-only for now.

We first covered how HeyClicky works before this comparison existed.

Choose it if: you want spoken, visual guidance inside a Mac app and may later hand off tasks.

Skip it if: you need Windows, searchable history, or a bring-your-own-key setup.

GhostLy: best for a small Mac overlay with your own model

GhostLy is closer to a universal shortcut for the AI accounts you already pay for. It reads what is on the current screen and clipboard, then lets you ask Claude, GPT or Gemini without moving context into another tab.

Its privacy page says prompts go directly to the selected AI provider and are not stored by GhostLy. The app processes screen text locally. The important caveat sits in the product copy: visual questions may still send a screen capture when needed. "Processed locally" therefore does not mean the model provider never sees the screen. It means GhostLy tries to turn the screen into smaller local text context first.

GhostLy costs $20 once, includes 20 free trial queries and was Mac-only when checked. It does not position itself as an agent that takes over the desktop.

Choose it if: you want current-screen answers, your own model account and no GhostLy subscription.

Skip it if: you want continuous memory or an agent that clicks and types.

Crade: best cross-platform option when the AI must act

Crade runs on Mac and Windows. It can answer from the active screen, then in Agent mode click, type and finish a task. You can connect an existing Claude or ChatGPT account or use Crade's built-in models.

The privacy policy is more useful than the homepage. A request can include your text, a snapshot of the active screen, the active app name and the browser domain. The screenshot is used to generate the reply and is not retained on Crade's servers afterwards. Text conversations are saved to your account so you can return to them. If you connect Claude, ChatGPT or Gemini, the screenshot and prompt go directly from your computer to that provider rather than through Crade.

That is a reasonable design, but it is still cloud processing. Crade's own policy tells users not to leave passwords, financial details, medical records or other regulated data visible while sending a request.

Choose it if: you need Mac and Windows support, current-screen help and optional desktop actions.

Skip it if: policy requires local-only inference or no saved conversation history.

AYO: best privacy design on paper, but still early

AYO takes a different route on Windows. Instead of uploading a screenshot, it says it reads screen text locally through accessibility APIs. The wake word, voice transcription and screen reading happen on the device. No screenshots, audio or raw screen text are saved. When you ask something, the question plus a short context summary goes to a cloud model on AWS Bedrock.

That separation is sensible. It reduces image exposure without pretending the answer is fully offline. AYO says local reading works without internet, but the reasoning step needs a connection.

The limitation is availability. AYO was still an early-access waitlist for Windows 10 and 11, English only, with final paid pricing not published. The desktop app is proprietary rather than open source. The vendor says free will remain available, but that is a future plan, not a current purchasing option.

Choose it if: you use Windows, like voice input and can evaluate an early product.

Skip it if: you need something available to a whole team now, macOS support or independently auditable application code.

Screenpipe: best when you need memory, not a second pair of eyes

Screenpipe is the outlier because it is designed to remember. It continuously captures screen changes and audio, extracts app text through accessibility APIs with OCR as a fallback, and stores the resulting text, transcripts and compressed screenshots in a local database. You can then ask what happened in a meeting, find the message where a price was discussed, or give an approved workflow the context it needs.

It supports Mac, Windows and Linux. Local models can keep the full flow on the machine. Connect a cloud model or sync service and selected context can leave the device, so deployment still needs a deliberate policy.

Screenpipe is source-available, not open source under a permissive licence. Its June 2026 licence change kept personal, educational and evaluation use free while requiring a paid licence for commercial use of the source. That distinction matters if a company plans to build on it rather than just subscribe to the signed app.

Choose it if: the question is "where did I see that?" or "what happened last week?"

Skip it if: you only want a lightweight current-screen helper. Continuous memory creates a larger archive to secure and govern.

Windows Recall: best built-in retrieval on a supported PC

Recall is not a general AI assistant, which is exactly why it belongs in the comparison. Search results often mix screen helpers with screen memory. Recall only solves the second job.

On a supported Copilot+ PC, Recall can save snapshots every few seconds and when the active window changes. It is opt-in, can be paused, supports app and website filters, and keeps snapshots and its search index encrypted on the device. Windows Hello protects access. Microsoft says the snapshots are not shared with Microsoft or third parties by default.

Recall's Click to Do layer can copy, rewrite, summarize or send selected content into another app. It is not the same as Crade or HeyClicky completing an open-ended desktop task.

Choose it if: you already own a compatible Windows PC and mainly need to retrieve past activity.

Skip it if: you need Mac or Linux, a conversational tutor, or an automation system.

The privacy check before installing any of them

Screen access is a powerful permission. A good product page is not enough for a company rollout.

Run one harmless workflow first. Use a test account, a document with fake customer data and an app that contains nothing confidential. Then check:

1. Whether capture starts only after the trigger you expect.
2. Whether screenshots or summaries appear in local app data.
3. Whether chat history can be disabled and deleted.
4. Which AI provider receives the request.
5. Whether password managers, private browser windows and chosen apps can be excluded.
6. Whether actions need confirmation before clicking, typing, sending or deleting.

For a team, decide the rule before installing the tool. Customer records, payroll, health data, legal work and production credentials should not become an accidental test set.

If the goal is to turn repeated on-screen work into instructions rather than keep a searchable diary, AgentHandover takes that narrower approach. For writing across apps, Aqua Voice reads the current context for dictation without trying to become a general screen agent.

Final verdict

For most Mac users asking this exact question, start with HeyClicky. It makes the current-screen value obvious and only looks after an explicit trigger.

For Mac users who already pay for a model and want a smaller utility, pick GhostLy. For a mixed Mac and Windows team that needs actions, evaluate Crade with test data before giving it a real workflow.

Choose Screenpipe or Windows Recall only when memory is the job. Continuous capture is not a better version of on-demand help. It is a different trade, more recall in exchange for a much larger record of the working day.

If this is a team workflow rather than a personal helper, we design AI automation around the data boundary first.

Sources

Product details and privacy claims were checked on August 16, 2026 against the official HeyClicky site and privacy policy, GhostLy product page and privacy policy, Crade product page and privacy policy, AYO product page, Screenpipe overview and licence update, and Microsoft's Recall guide and privacy controls.

Words worth knowing

Screen-aware: software that uses what is currently visible as context instead of relying only on what you type.

Accessibility API: the structured text, labels and controls an operating system exposes for screen readers. An AI can use it to understand an app without relying only on pixels.

OCR: software that reads text inside an image or screenshot.

Local inference: the AI model runs on your own computer. Local capture without local inference can still send a summary or image to a cloud model.

Continuous memory: a background record of screen or audio activity that can be searched later.