A companion that stays close
A compact floating card keeps the current instruction and target status visible. Expand it to ask about the step or start a new task.
A little guidance.
A lot more
possible.
Describe what you want to do. Desktop Guide opens a browser, reads your task window, and helps with the next step. Follow the green guide, or choose the red autonomous agent to click, scroll and type.
Windows 11 · x64 · Development preview
YOUR NEXT JOURNEY
Where to next?
New DelhiFROMJaipurTO
Search trains
Your task windowYou click and type
Desktop
Guide
YOUR NEXT STEP
Find your
next step.
Choose Search trains to see the available options.
Your visual guide.Find click target
Check screen
Check screen
Waiting for you
Illustrative product preview
Start with a prompt
Describe the goal. Gemini chooses a starting site and browser.
Bring help to your app
Press Ctrl+Alt+Space to prepare help with the window in front.
You make the move
Follow a drawn cursor and clear instructions. Your mouse stays yours.
Your screen.
Your pace.
From your first prompt to the next visible result, see how the current companion works.
Start with
your goal
EXAMPLE TASK
“Open Chrome and help me find a train ticket.”
In the app, starting sends your prompt for planning, then shares live screenshots of the task window for guidance.
Find your next journey.
FromNew Delhi
ToJaipur
Click here
Journey options are on screen.
Illustrative search result · No booking made
Your task windowVisual guide · You make the click
Illustrative demo · No screen access or AI requests
Desktop Guide
A goal is all you need to start.
Describe the task in your own words. Gemini chooses a starting site and opens a new browser window, or asks for clarification.
In the app, guidance comes from the current screenshot. You perform each action.
Already in the right app?
Bring it to the front, press the shortcut, describe your goal, then choose Help with this window. The shortcut alone shares no images.
Ctrl+Alt+Space
A helping hand.
You hold
the controls.
Live help stays close to the task, from the first instruction to a fresh check of the result.
See where to look
Guided mode draws a cursor and click ring for you to follow. Autonomous mode animates the controlled cursor as it moves to the target. Both respect your reduced-motion preference.
Find it. Then check it.
Find click target retries the current instruction. Check screen checks what happened after your action. Loading pages wait for a visible change before the next step.
Stop means stop
Pause before sensitive input. Stop ends capture and cancels pending work locally. Resume asks for fresh consent to observe your window.
Guidance from your screen
The current build uses live screenshots to choose instructions and targets. Uncertain, covered or stale targets are withheld.
Keep your place through a hiccup
Lost responses are recovered first. Temporary analysis failures get up to 12 fresh-image retries, with visible progress. Resume reconnects in the same mode and plans from the current screen.
Your next step
starts here.
Download Desktop Guide for Windows
and bring a guide to your next task.
Windows development preview
Windows 11 · x64 · Portable ZIP
Unsigned development build. An invited pilot account and internet connection are required. The latest version uses live cloud guidance, optional autonomous actions, voice prompts and phone control.
Every release. One place.
Choose a version, read what changed, and verify your download.
Release notes
The blob, thinking/action dock, cursor, outlines and panels are now visible in desktop recordings and screen sharing. Share the entire screen or monitor to include these separate floating windows. Agent screenshots still isolate or locally mask the assistant UI. Opening and closing the deck uses a smooth slide/fade, respects Reduced motion and handles rapid toggles. Visible Exit app controls stop the task and quit completely. Retains r20 verified task history, meeting-to-WhatsApp Desktop handoff when named in your prompt, popup-cycle replanning and 12 recovery retries. Focused deterministic, headless UI and native synthetic capture checks passed without model calls or real input. Signed-in Meet/WhatsApp and individual third-party recorders remain untested. Requires invited pilot account and internet. Unsigned development build.
SHA-256 checksum7d4860f0b938b7d23dbda88caf8e1c7b05593ddf5edcaef7386cf46414584209114.5 MB · Unsigned development build · Windows 11 x64
Release notes
Keeps the whole task and bounded verified history across screenshot turns and Resume. After creating and copying a meeting link, the agent can open installed WhatsApp Desktop when explicitly named in your prompt and continue searching for the requested chat. Detects popup action cycles even when pixels change; recovery is limited to 12 retries. App handoff preserves the task action budget and expiry, rejects late responses after Stop, and does not reset the session or goal. Live Gemini synthetic image checks passed Copy link, Open WhatsApp and chat search. Signed-in Google Meet and WhatsApp together remain untested. Requires invited pilot account and internet. Unsigned development build.
SHA-256 checksumbf10fd8f7ef424ba3b47b64106f962dbec58076f40fcbf2821b9e2cd485ba270114.5 MB · Unsigned development build · Windows 11 x64
Release notes
Fixes the autonomous cursor overlay covering the button before a click. Cursor travel remains animated; overlays are cleared before input and the click ring appears after dispatch. Guided marker requests cannot interrupt Autonomous mode. A separate-process Windows test reproduced the covered target and then passed one real click with this fix. Retains r18 meeting-link creation and recovery, Guided mode, voice, phone control, Resume and Stop. Requires an invited pilot account and internet. Unsigned development build; the signed-in Google Meet workflow itself remains untested.
SHA-256 checksum7a15c154b23fdd18d8701079c261be87cbc61879f0cada4649682e4403fe0574114.5 MB · Unsigned development build · Windows 11 x64
Release notes
Fixes autonomous clicks being deferred by small page changes and recovery asking you to complete a click that was never sent. Hover changes trigger fresh screenshot grounding and another bounded attempt. Meeting-link creation is treated as an ordinary authorized task action. A disposable Windows test against the deployed Gemini service completed New meeting, Create a meeting for later, and verified the resulting link with two real clicks and no manual handoff. Guided mode, voice, phone control, Resume and Stop are retained. Requires an invited pilot account and internet. Unsigned development build; the signed-in Google Meet workflow itself remains untested.
SHA-256 checksum1d50a4dc011e9f1e698863779557e47bc0f94c4c53d92b003b69bce5dddfda34114.5 MB · Unsigned development build · Windows 11 x64
Release notes
Guided and Autonomous modes, green/red animated companion, voice prompts and locally enabled phone control. Up to three retries for temporary analysis failures, fresh-screen Resume in the original mode, immediate local Stop, visible thinking/retry progress and an animated controlled cursor with click rings. Live screenshot decisions only. Real Windows test against the deployed Gemini service passed retry, pause/resume, verified click and Stop during cursor movement. Requires an invited pilot account and internet. Unsigned development build; full IRCTC and broader real-site workflows remain under qualification.
SHA-256 checksumb26c532da5a95e934244fa6811f567747cf026fa03346471a935c9f8910c1643114.5 MB · Unsigned development build · Windows 11 x64
Release notes
This archived pilot: describe a goal to let Gemini choose a browser and starting site, or use Ctrl+Alt+Space for an existing window. Live screenshot guidance now uses a compact companion, drawn cursor/ring/callout, Find click target, Check screen, loading-state handling and interrupted-response recovery. Screenshot-only guidance; offline walkthroughs and browser control inspection are disabled. You click and type; autonomous execution is not included. Requires an invited pilot account and internet. Unsigned; full native workflows, voice and accessibility remain under qualification.
SHA-256 checksumd5fcace83287d657fade82399192b39df85ae345bf2c49c0768e56173ae11514114.5 MB · Unsigned development build · Windows 11 x64
Release notes
Windows development pilot with an offline walkthrough, invite-only cloud sign-in, guided lesson flows, optional pointing, and separately approved single-action demonstrations. Includes the .NET runtime. This portable ZIP is unsigned; broader real-workflow and accessibility qualification is still in progress.
SHA-256 checksume73fa1eff9691cad19c9b805c69c83bb09248373d69d9513f1d46533adffd87e91.9 MB · Unsigned development build · Windows 11 x64
No stable releases yet. The development preview is available under All releases.
7 available releases
How do I install Desktop Guide?
Download the ZIP, extract all files to a folder you can write to, and open DesktopGuide.App.exe. Windows 11 x64 is required. The .NET runtime is included. This preview is unsigned, so Windows may display a security warning. No installer or automatic updater is included. To update, close the app from its tray menu and extract the latest ZIP into a new folder. Older extracted copies do not update automatically.
Do I need an account?
Yes. The latest build needs an invited pilot account and an internet connection for live guidance. Sign in with your invited email and the permanent password you set. Downloading the ZIP does not create an account. Offline walkthroughs have been removed from the current version; the older archived pilot has different behavior.
What can the development preview do?
Start from a prompt or the current-window shortcut, follow live screenshot guidance in the floating companion, retry a target, check your result, or allow autonomous actions. Speak a prompt, review its text, or connect the phone controller from desktop settings. Chrome is the default for new browser tasks; name Edge or Firefox in your prompt if preferred. Full real-application workflows, native overlay placement, voice and accessibility are still being qualified. This is an unsigned development pilot; the website demo is illustrative.
What does the app see?
Starting a task sends your prompt for cloud planning, then shares live screenshots of the task window for guidance. The current build uses screenshots only; DOM and Windows accessibility inspection are disabled. An optional manual source/preview flow supports privacy masks. The app is designed not to persist raw learner screenshots or audio. This website never captures your desktop.
Will it click, type or deploy for me?
Choose Guided to follow the cursor yourself, or Autonomous to let the agent click, scroll and type in your selected task window. Autonomous tasks need your local permission and run for up to 30 actions or 10 minutes. You handle login, CAPTCHA, billing and final submissions. Moving your mouse or typing pauses the agent.
What if the target is missing or the page is loading?
The companion shows an animated analysis indicator and retries temporary analysis failures up to three times with fresh images. Autonomous mode also rechecks blocked targets before pausing. Resume reconnects, reads the current screen and keeps your chosen mode. Stop cancels locally and clears permissions immediately. Bring the task window back to the front if focus is needed.
How do I pause, stop or exit?
Pause is available in the expanded companion. Stop is available on the dock, in the tray and through Ctrl+Alt+Shift+S when the shortcut is registered. Stop ends capture and cancels pending work; resuming requires fresh observation consent. Closing the main panel leaves the companion running. Choose Exit from the tray to quit.