Gemini Spark vs OpenClaw: Google's agent, or one of your own?
Google's always-on agent against the open-source one you run yourself: the computer, models, chat apps, what each reaches, data and who can have it.
Both are an AI agent with a computer of its own that keeps working while you’re away, runs jobs on a schedule, follows skills written as SKILL.md files and reaches other apps through MCP. Gemini Spark is Google’s own ready-made service inside the Gemini app: deepest in Gmail, Calendar and Drive, Gemini models only, $19.99 a month with Google AI Pro in the US, for personal Google accounts outside the European Economic Area, the UK, Switzerland and Nigeria, with your activity used to train Google’s models. OpenClaw is free, open-source software you run on a computer you choose, with any model you like, in the chat apps you already use. The choice is between an agent that lives inside one company’s apps and an agent of your own that you can move.
The same idea, built two ways
Gemini Spark, announced at Google I/O on May 19, 2026, is a mode of the Gemini app: you give it tasks, schedules and skills, and it works through your Google account on a machine Google runs. OpenClaw is the open-source version of the same ambition, MIT-licensed, with its memory in plain files and a home on a computer you control. One arrives working inside the account you already have; the other arrives as software plus a few decisions about where it lives and which model it uses.
The computer underneath
Google says Spark “runs on dedicated virtual machines on Google Cloud,” with a remote browser that keeps your sign-ins and a remote computer that runs code and keeps files, and publishes nothing else about it: not its size, its operating system, where it is, or whether you could reach it yourself. Its files and cookies stay until you delete them in Spark’s settings, and switching Spark off deletes them; Google’s pages disagree on whether a schedule runs while all your devices are off, which the explainer quotes. OpenClaw’s computer is whichever you give it, a laptop, a Mac mini or a server, so you know what it is and can open it, back it up or move it; the boundaries around it are yours to draw, which means drawing them well is your job too.
Models and the meter
Spark runs on Gemini: Gemini 3.5 at launch and Gemini 3.7 Flash since August 13, with no picker. It shares the Gemini app’s usage limit, which refreshes every five hours up to a weekly cap, and Google publishes no figure for how much agent work a plan buys; Ultra, at $99.99 or $199.99 a month, raises it five or twenty times. OpenClaw takes any model with an API, and you can change it per task, run a cheap model for routine jobs, or use a ChatGPT subscription you already pay for. The software costs nothing; what it costs to run is the machine plus the model, both itemized.
Where you talk to it
Spark lives in the Gemini app on the web, Android, iPhone and Mac, and in Gemini Live’s voice chats; the Windows app has none, though the web app works there, and texting or emailing Spark, promised at I/O for “soon,” isn’t documented as available. OpenClaw lives in the messengers you already open every day: WhatsApp, Telegram, Discord, Slack, Signal and more, so the agent is a contact in your messages rather than a place you go to, and it can message you first when something needs you.
What each can reach
Spark’s advantage is Google’s own apps. It reads and acts in Gmail, Calendar, Drive, Docs, Sheets, Keep and Tasks out of the box once you connect them, plus about 40 other connected apps, most of them US only, your own Chrome with its saved passwords if you allow it, and in the US your own MCP servers. OpenClaw reaches what you connect: Gmail and calendars through their own access, any MCP server, a browser on its machine, the files and command line of the computer it runs on, and skills in the same SKILL.md shape. More setup, and no single company’s apps favoured over another’s.
Safety and what the company keeps
Both have the same unsolved problem: prompt injection, where a web page or an email carries instructions meant for the agent. Google’s help page lists the outcomes (your email sent to an outside service, private information posted publicly) and says its safeguards “don’t guarantee protection against all risks”; OpenClaw’s own security docs say the same of any untrusted content the agent reads. Spark answers with confirmations before sending, buying, submitting or changing data, and hands payments back to you. OpenClaw answers with what you configure: who may message it, which commands need your approval, what the machine holds.
Where they differ most is data. Spark requires Keep Activity on, and under that setting, per the Gemini Apps Privacy Hub, Google uses your activity to improve its services, “including training generative AI models,” with human reviewers; reviewed chats are kept up to three years. With OpenClaw, your conversations and memory stay on its computer, and what reaches a model is governed by the provider you picked and its terms, which you can change.
Who can have it
Spark needs a Google AI Pro or Ultra plan, a personal Google account (work and school accounts are excluded “for now”), an age of 18, and a country outside the European Economic Area, the UK, Switzerland and Nigeria. OpenClaw is software: anyone can run it anywhere, under any account, for themselves or for a business, subject only to the model provider’s terms.
The choice
If your life runs on Gmail and Google Calendar, you already pay for Google AI Pro, and you want something that works the day you switch it on, Spark is the least-effort version of this idea, and Google’s apps are where it is strongest. Choose OpenClaw when the agent is meant to be yours: your model, your chat apps, your files on a machine you can name, in any country, and nothing deciding what it is allowed to connect to. The running of that machine is the part you can hand off: hosted OpenClaw puts it on a server someone else keeps. Everpod’s version puts your agent on its own private cloud computer, set up for you with no technical knowledge needed; your pod arrives secured, backed up, and private to you. First month half price, then $29 a month, cancel anytime. Your plan covers up to $10 of model usage a month.