GuidesGetting started

What is Gemini Spark? Google's always-on AI agent, explained

Google's 24/7 personal agent in the Gemini app: what it runs on, what it can reach, when it asks first, what Google keeps, who can get it and what it costs.

October 9, 2026The Everpod team
The short answer

Gemini Spark is Google’s always-on personal AI agent, announced at Google I/O on May 19, 2026. You give it a task or a schedule in the Gemini app, and it carries the work out on Google’s cloud, through your Gmail, Calendar, Drive and other connected apps and a browser, “even if your phone and laptop are turned off,” asking you before it sends, buys, submits a form or changes your data. It comes with Google AI Pro ($19.99 a month in the US) and Ultra ($99.99 or $199.99), for people 18 and over with a personal Google account, everywhere the Gemini app is offered except the European Economic Area, the UK, Switzerland and Nigeria.

It is a mode of the Gemini app, not a separate app, and it is not the business “Gemini agent” Google Cloud announced on October 8, 2026.

Which Gemini agent is this?

The computer behind it

Sundar Pichai’s I/O post gives the one line Google has published about the machine: “It runs on dedicated virtual machines on Google Cloud. And it’s 24/7 so you don’t need to keep your laptop open.” The help pages describe two pieces of it: a remote browser, “a separate, new browser instance, not your local browser in the cloud,” which keeps your sign-ins between sessions, and a remote computer where Spark runs code and keeps files. Google doesn’t say what that computer is (its operating system, size or tools), whether each person gets their own, or where it runs. Google’s pages disagree on one point that matters for schedules: the schedules page says Spark “keeps working even if your device is turned off,” while the main help page’s section on turning Spark off says “Schedules will not run as long as Spark is off or if your device is off.” Until Google settles it, don’t rely on a schedule running while every device of yours is off. At launch Spark ran on Gemini 3.5 and Google’s Antigravity agent harness; since August 13 it uses Gemini 3.7 Flash, the latest model Google has named for it. There is no model picker.

Spark is built from three things, which Google calls “the what, when, and how”: a task is a goal you give it; a schedule runs a task on its own, at a time (once, hourly, up to yearly), when a Gmail filter matches a new email, or when a news or market topic moves; and a skill is reusable instructions. Run times are “approximate,” and a schedule is “locked to the timezone of the location where it was created.” Skills are written in chat or uploaded as a SKILL.md file or a .zip holding one, with a lowercase, hyphenated name, up to 100 MB, and scripts that don’t reach the internet; you can download a skill as a .zip.

What it can reach, and when it asks first

Through Google’s own services it can read and act in Gmail (search, drafts, replies, labels), Calendar (RSVPs, new and changed events), Drive, Docs, Sheets and Slides, Keep, Tasks, Contacts, Photos in the US, YouTube, and Search, Maps, Flights and Hotels. About 40 connected apps have a Spark mode, most of them US only, among them Canva, Instacart, OpenTable, Dropbox, Zillow, Linear and Ticketmaster. In the US, in English, you can connect your own app by giving the web app an MCP server’s address. On a Mac it can work in the folders you add, and in desktop Chrome it can use your own browser, “including sites you’re signed into,” and with your permission the passwords Chrome has saved. Google’s apps are “turned off by default”: you connect each one.

The help page says Spark is designed to ask for your review before “sending communications, modifying your data, making purchases, and submitting web forms,” and before signing in with Google. It names its own exceptions: bulk actions in Google Tasks run without asking, a Mac file change you request is made “by default,” and “if a schedule runs when you are offline, you may not be able to stop Gemini from completing an unintended action.” Payments go back to you: Spark asks you to “take control” to enter a password or card details, and Google says not to paste either into a task. Google announced in May that its Agent Payments Protocol, which lets you set what an agent may buy and how much it can spend, would come to Google products “starting with Gemini Spark”; no help page describes it as available yet.

Google is direct about what can go wrong. Spark is “an experimental feature in early development,” and the help page describes prompt injection with examples: a web page, email or document carrying hidden instructions that make the agent post your private information publicly, send your email to an outside service, or install compromised software. Its safeguards, it says, “don’t guarantee protection against all risks,” and “your active supervision is the most important way to protect against risk while using Spark.” Chrome’s page adds: “You’re responsible for Gemini’s actions during a task, including mistakes and unexpected results like purchases.”

What Google keeps

Spark needs Keep Activity switched on, and under the Gemini Apps Privacy Hub that setting means Google “uses your activity to provide, develop, and improve its services (including training generative AI models),” with human reviewers. Activity is deleted after 18 months unless you choose 3 or 36 months, or never; chats a reviewer has seen are kept for up to three years even after you delete them. So the usual way to keep Gemini chats out of training, switching Keep Activity off, isn’t open to someone using Spark. Google says your Gemini chats are not being used to show you ads.

The remote browser’s cookies and the remote computer’s files stay until you delete them in Spark’s settings, and switching Spark off deletes both; deleting a task removes the task, its activity and its schedules but not files it changed in Drive or the browser data. Google’s export tool covers your Google data, and skills download one by one; nothing describes exporting tasks, schedules or the remote computer’s files.

Who can use it, and what it costs

Requirements, from the help page as of October 9, 2026: 18 or over, a personal Google account (“for now, this feature isn’t available if you sign in with a work or school Google Account”), a Google AI Pro or Ultra plan, and Keep Activity on, anywhere the Gemini app is offered except the EEA, the UK, Switzerland and Nigeria. Google’s own pages count those places differently, from “over 160 countries” to “available globally.” It opened in beta to Ultra subscribers in the US in May, reached the Mac on June 30, Ultra subscribers in every Gemini language outside those four places on July 14, Pro subscribers in the US on July 16 and Pro in 160 more countries on July 30. Google’s Windows app has no Spark, though the web app works in any browser, and texting or emailing Spark, promised in May for “soon,” isn’t documented as available.

Spark has no meter of its own: it draws on the Gemini app’s compute-based limit, which “refreshes every 5 hours until you reach your weekly limit,” and Google publishes no figure for how much Spark work a plan buys. The fixed caps are 15 tasks running at once, 50 active schedules, and, for browsing in your own Chrome, 20 requests a day on Pro and 200 on Ultra. Cancel the plan and running tasks finish while schedules pause, and nothing is deleted.

Your own open-source AI agent, set up for you.

Everpod runs OpenClaw on a private, always-on computer of its own: set up, secured and backed up, with model usage included. You name your agent, and say hello about fifteen minutes later.

Create your agent

First month half price, then $29/mo · model usage included · cancel anytime

Wondering what you’d do with one? See what a cloud agent can do