OpenClaw and Google Spark split on where agents live

OpenClaw crossed 300,000 GitHub stars in April and became one of GitHub’s fastest-growing repositories, according to the article. Around the same time, Google used its I/O event to launch Gemini Spark, a 24/7 personal agent built on Gemini 3.5 Flash and tied into Google’s Antigravity agent stack.

The headline difference is where the software runs. OpenClaw’s appeal is that it lives on hardware the user owns, such as a Mac mini sitting on a shelf and drawing about seven watts while it runs in the background. Google’s Spark takes the opposite approach: it runs on virtual machines in Google Cloud, and the user never sees the machine.

According to the article, both systems are meant to do similar work. They can watch an inbox, draft a status update, browse the web, and handle recurring tasks. Both are also moving toward Model Context Protocol, or MCP, for connecting tools, though the article says their implementations are at different stages of maturity.

The deeper difference is control. With OpenClaw, the user buys the hardware and sets up the system themselves. The article describes the setup as buying a Mac mini, keeping it awake, installing a daemon, configuring Tailscale, and rotating keys when they expire.

That extra work buys ownership of the environment, including credentials and workflows, depending on how the user wires models and integrations. The article argues that this matters because the substrate - the underlying place where the agent runs - determines who holds the context, who can see credentials, and who can change the rules later.

Google’s Spark is easier to adopt because it sits inside the company’s own products. The article says it is already inside Gmail, Docs, and Sheets, which gives it broad access without manual wiring. Google also plans to let users text and email the agent directly, so it can keep working even when a laptop is closed.

That convenience comes with tradeoffs. The article says a local agent with shell, browser, and inbox access can be risky if it is misconfigured, and that Chinese regulators have already flagged that concern with OpenClaw. It also says the self-hosted model is not about nostalgia for running servers; it is about wanting an agent intimate enough to manage daily work but still answerable to the user.

The article places the market in two camps: a hosted tier where Google, and later OpenAI, own the runtime and context, and a self-hosted tier for developers and privacy-sensitive users who want credentials on their own hardware. It argues that managed services usually win for most users, but that OpenClaw can still hold a durable base among people who care enough about control to do the setup work.

The piece closes by framing the real question for developers as whether they are comfortable with Google holding the keys to the agent that runs their life.