Google brought Gemini Spark to macOS on Wednesday, adding a desktop build of its agentic assistant to the existing Gemini app and pushing directly into the desktop-agent category defined by Claude Desktop, Microsoft Copilot, and OpenClaw. Gemini Spark on Mac is in beta and restricted to Google AI Ultra subscribers in the U.S., a tight rollout that keeps the launch inside Google's paying power-user base. Alongside the Mac build, Google added seven new app integrations and turned on custom Model Context Protocol support.
The Mac version lets Spark reach into local files — sorting and organizing them, or using them as source material for a new Google Workspace doc or spreadsheet. Google's own example: point Spark at a folder of invoices and have it produce a budgeting worksheet. That local-file access is the feature that puts Spark on the same competitive tier as Claude Desktop, which has anchored the desktop-agent segment since Anthropic shipped it, and Microsoft's Copilot, which sits inside the Windows and Office surface.
The new integrations are the bigger unlock in day-to-day use. Spark now connects to Google Tasks and Google Keep — the Keep gap was a real one, since short lists and quick notes never belonged in Google Docs, and its absence at Spark's launch last month was the most obvious rough edge. Third-party integrations now cover Canva, Dropbox, Instacart, OpenTable, and Zillow Rentals, which is what actually turns Spark into a booking-and-ordering agent rather than a chat window with file access.
Key facts
- 01Gemini Spark launched on macOS in beta on Wednesday, limited to Google AI Ultra subscribers in the U.S.
- 02Spark added integrations with Google Tasks, Google Keep, Canva, Dropbox, Instacart, OpenTable, and Zillow Rentals.
- 03Google is rolling out custom Model Context Protocol support so users can wire third-party apps directly into Spark.
- 04The Mac release puts Spark head-to-head with Claude Desktop, Microsoft Copilot, and OpenClaw on the desktop-agent tier.
- 05Cross-device handoff — assigning multi-step tasks from a phone to the desktop agent — is promised but not shipping at launch.
Spark can also now track topics and react to events in real time. Google is pitching that for sports scores, stock movements, and breaking news, but the same primitive works for social media, blogs, online shopping, and weather. In practice, a persistent watch loop is what separates an agent from a chatbot — the model doesn't just answer when prompted, it acts when a condition it's watching changes.
Custom Model Context Protocol support is the developer-facing piece and arguably the most consequential. MCP, the open protocol originally pushed by Anthropic and now adopted across the frontier labs, lets any compliant tool plug into any compliant model client. By supporting custom MCP servers in Spark, Google is conceding that a closed integrations catalog isn't going to work — users want to wire their own tools in, and the industry has settled on MCP as the way to do it.
The competitive backdrop matters. Claude Desktop set the template for a desktop-resident AI that can see local files and drive local apps. Microsoft has been layering Copilot deeper into Windows on the same premise. OpenClaw has been chipping away at the open-source end of the same problem. Google is arriving with the advantage of an existing Gemini install base and the Workspace hooks that already sit on most of its Ultra subscribers' machines.
The rough edge on the Mac build is cross-device handoff. Google says users will "soon" be able to assign multi-step tasks to Spark from their phones — calling up the desktop agent to pull information from a file on the Mac, for instance — but that isn't shipping at launch. It's the kind of feature that reads well in a keynote and takes a full quarter of engineering to make reliable, so "soon" should be read as "soon-ish."
Google's positioning here is that Spark is the assistant that already knows your Gmail, your Calendar, your Docs, and — as of this week — your Keep notes and your Tasks list, and that runs on your Mac with access to your local files and a growing catalog of third-party apps. That's a defensible pitch against Claude Desktop's cleaner agent loop and against Copilot's tighter Windows integration, and it lands specifically at the users paying $249 a month for AI Ultra.
The MCP support is the piece worth watching. Every frontier client that opens up to custom MCP servers pushes the industry closer to a world where the model client is a commodity and the value sits in the tools you connect to it. Google shipping MCP into Spark on the same day it launches on Mac is a bet that being the most open desktop agent is worth more than being the most locked-down one — and it's the right bet, because the users deciding between Spark and Claude Desktop this quarter are exactly the users who will judge the winner on how well their existing tool stack plugs in.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
