Vivold Consulting
Product & Tech Updates

Gemini Spark lands on the Mac: Google's 24/7 agent starts working your local files

Google's agentic assistant gains desktop file access, real-time tracking, Keep/Tasks and MCP support - the desktop agent race is on

Key Insights

Gemini Spark, Google's 24/7 agentic assistant, is now available on Mac (beta, US-only, Google AI Ultra subscribers), where it can work directly with files on the computer - sorting and organising them, or turning a folder of invoices into a budgeting worksheet in Google Workspace. The update adds long-requested Google Tasks and Keep integrations plus third-party hooks into Canva, Dropbox, Instacart, OpenTable, and Zillow Rentals, real-time tracking of topics like stocks and breaking news, and - notably for builders - custom MCP support for wiring in your own apps. It puts Spark in direct competition with Claude Desktop, Microsoft Copilot, and OpenClaw for the desktop, where the real productivity (and governance) questions live.

Stay Updated

Get the latest insights delivered to your inbox

The agent moves in with your files

Google's Gemini Spark - the always-on agentic assistant it launched only about a month ago - has arrived on macOS, folded into the existing Gemini desktop app. The significance is location: on the desktop, Spark can finally work with the files on your computer, sorting and organising them or using them as source material for new Google Workspace documents and spreadsheets. Google's own example is telling for anyone who runs a business: point Spark at the invoices sitting on your Mac and get back a budgeting worksheet. Multi-step phone-to-desktop handoffs - asking the mobile app to pull information from a file on your Mac - are promised soon. For now the beta is limited to Google AI Ultra subscribers in the US, a narrow gate that signals early days.

Filling the gaps, fast

The update reads like a direct response to early-user complaints. Reviewers had specifically flagged the missing Google Keep integration - short lists and quick notes were being forced into full Google Docs - and Google has now added both Keep and Tasks. Third-party reach expands too: Canva, Dropbox, Instacart, OpenTable, and Zillow Rentals join the roster, letting Spark design flyers, order the weekly groceries, book restaurant tables, or schedule apartment tours. A new real-time tracking capability lets the agent monitor topics and react to events as they happen - sports scores, stock movements, breaking news, social media, even online shopping and weather. And in the most strategically interesting line of the announcement, Google is rolling out support for custom Model Context Protocol (MCP) connections, letting users wire their own preferred apps directly into Spark.

The desktop is the new front line

The macOS launch is explicitly a competitive move against Claude Desktop, Microsoft's Copilot, and OpenClaw - the class of agents that live where work actually happens, among local files and installed applications, rather than in a browser tab. The desktop is where an agent's usefulness compounds (it can see the messy reality of your file system) and where the governance stakes rise for exactly the same reason.

How to put desktop agents to work without getting burned

  • The MCP support is the strategic headline for anyone building internal tools: Google embracing the same connector protocol as the Anthropic ecosystem means integrations you build once increasingly travel across assistants. If you are investing in agent plumbing, invest in MCP connectors, not per-vendor plugins.
  • A high-yield, low-risk pilot for operations teams: file-heavy chores like invoice collation, folder hygiene, and turning document piles into structured spreadsheets are precisely what this release automates. Measure hours saved on one workflow before rolling wider.
  • Do the governance work before the rollout: a 24/7 agent with local file access and real-time web monitoring is a new data-boundary question. Decide which machines, which folders, and which accounts an agent may touch - and write it down - while adoption is still small enough to shape.
  • Calibrate expectations on enterprise readiness: beta status, US-only availability, and the premium Ultra subscription gate mean this is a signal of direction, not a fleet-deployment candidate. The direction, though, is unambiguous - every major vendor now wants an agent resident on your desktop, so choosing your default (and your data rules) is a 2026 decision, not a someday one.

More in Product & Tech Updates

All Product & Tech Updates stories

Claude gets a 'Reflect' dashboard: Spotify Wrapped for your AI habit - and a masterclass in retention design

Anthropic launched Reflect, a beta dashboard (Free, Pro, and Max users with Memory on) that visualises Claude usage over 1-12 months - top topics, task types, peak hours - and coaches you via its 4D AI Fluency Framework (delegation, description, discernment, diligence), suggesting features like Projects or custom skills based on your patterns. It ships wellbeing controls (quiet hours, break nudges, reflection prompts) built with MIT Media Lab and Boston Children's Hospital experts, and excludes health-linked conversations entirely. TechCrunch's sharp read: beneath the mindfulness framing, Reflect showcases how much of your work runs through Claude - a retention play as much as a wellness one.

L'Oreal brings Maybelline virtual try-on to ChatGPT

L'Oreal has announced a wide-ranging collaboration with OpenAI, unveiled at VivaTech 2026, that brings Maybelline's virtual makeup try-on directly into ChatGPT via L'Oreal's ModiFace AR technology. The deal spans consumer shopping tools, product discovery for brands like Lancome and Kerastase, advertising pilots (SkinCeuticals, CeraVe, Garnier), and R&D - including using OpenAI's GPT-Rosalind life-sciences model for skin-microbiome research. It lands as OpenAI reports ChatGPT at more than 900 million weekly users.

Sakana AI launches Fugu, an orchestration model that matches the frontier

Sakana AI has launched Fugu and Fugu Ultra, a multi-agent orchestration system delivered as a single foundation model - Fugu is itself an LLM trained to route tasks across a swappable pool of the world's best models (and recursively to itself) via one OpenAI-compatible API. Sakana says Fugu Ultra matches frontier models like Anthropic's Fable 5 and Mythos Preview on demanding engineering, science, and reasoning benchmarks, while pitching the approach as an AI-sovereignty hedge: if one provider's access disappears, as with Anthropic's recently export-controlled models, Fugu reroutes around it. It is generally available today through subscription and pay-as-you-go tiers.