The reveal of Gemini Spark on the I/O 2026 stage had left an unanswered question: what can a personal agent that runs 24/7 on a Google Cloud VM really do if it is locked inside Workspace? The answer landed in the days after, and it is the most underrated piece of the keynote: Spark adopts the Model Context Protocol (MCP), the open standard born at Anthropic, as its primary channel for talking to third-party services.
The launch partners
The first live integrations in May are Canva (design), Instacart (groceries) and OpenTable (restaurant bookings, with Reserve with Google closing the loop). In the following weeks Adobe, Asana, Dropbox, Lyft, Uber, Zillow, Zocdoc and Shopify come online. Google described the first wave as 30-plus services, with more partners on the way.
The choice of MCP is the story inside the story. It means Google is giving up on a proprietary standard for its flagship agent and aligning with the same protocol that ChatGPT and Claude already speak. For connector developers it is a huge win: one MCP server can serve multiple agents without rewrites. For Google it is an acknowledgement that the value no longer sits in protecting your framework — it sits in having the widest connection network inside the agent the consumer will open first.
What changes for users
The operating model is simple: a user asks Spark to book a table for four nearby on Saturday night, and Spark completes the action on OpenTable without the user opening the app. Same logic for groceries on Instacart or a ride on Lyft. The difference from the previous generation of voice assistants is that Spark does not only run when the user is in front of a device: it runs all the time, because it lives on a Google Cloud VM, and it can complete multi-step flows in the background.
What is worth watching
The critical point is consent: every new permission Spark requests from a third-party service is a new attack surface. Google promises granular controls, but the history of mobile permissions suggests that most users grant everything at the first prompt. The second issue is cost transparency: Spark is bundled with AI Ultra in the United States from the week after the keynote, but third-party actions can trigger indirect costs (the groceries you actually buy on Instacart, the booking you pay for on OpenTable). The line between assistance and purchase is not always clean.