September 16, 2026
Google is opening early access to a new Model Context Protocol (MCP) server for Google Home, according to TechCrunch. That's a fancy way of saying Claude, ChatGPT, and other AI agents can now talk directly to your connected devices -- turning off lights, pulling up camera summaries, checking whether the garage door is open -- using plain language instead of a dedicated app.
The consumer framing is cute: ask your agent to lock the doors before bed. But the real story is architectural. Google just admitted that the interface layer for smart devices is moving away from apps and toward agents that sit on top of them. That's the same shift happening in business software right now, and it's worth taking seriously even if you've never owned a smart thermostat. Once an AI agent can operate a physical device on your say-so, the next question is what else it should be trusted to touch -- your calendar, your CRM, your payroll system. Google is running that experiment in the home first because the stakes of a misfire are lower than, say, an agent misreading a customer contract. I'd expect the permission and audit-log questions raised here to resurface almost immediately in enterprise contexts, and companies evaluating any agent-driven workflow should be asking now who's watching what the agent actually does, not just what it's capable of doing. That's the same reasoning behind role-based permissions in any serious business platform: capability without a permission layer is a liability, not a feature.
Anthropic is merging Claude's chat interface with Cowork, its work-focused product, into a single experience, TechCrunch reports -- rolling out first to Pro and Max subscribers. On its face this sounds like a UI cleanup. It isn't. Splitting 'ask a question' from 'do the work' was always an artificial distinction, and every major AI vendor has been converging toward one continuous surface where a conversation turns into a task turns into a finished deliverable without you switching windows.
For businesses already juggling five or six SaaS logins, this convergence matters more than the individual feature list. The fewer places work happens, the fewer seams where things get lost -- a dropped file, a stale spreadsheet, a task nobody owns. That's the exact problem a unified platform is supposed to solve, whether it's an AI chat tool or an operations stack. If you're currently stitching together a project tool, a CRM, and a document editor and calling it a system, Anthropic's move is a reminder that the market is voting with its feet against fragmentation -- a case we've made before about replacing scattered point tools with one platform instead of ten. The honest caveat: consolidation only helps if the merged tool is actually good at both halves of the job. A chat interface bolted onto a work tool that neither answers questions well nor gets tasks done cleanly is worse than two separate mediocre tools. Early access being Pro/Max only means most of us will be reading reviews before judging.
At TechCrunch Disrupt 2026, Nvidia's Les Karpas is expected to make the case for why robotics hasn't had its 'ChatGPT moment' yet -- the point where the technology stops being a demo and starts being something people rely on daily. It's a useful contrast to the other two stories: software agents are already inside your smart speaker and your work chat, while physical robots are still stuck explaining why they're not. The gap isn't hardware, mostly -- it's the same trust and reliability problem software agents are only now starting to solve at scale. Businesses evaluating any AI agent, physical or digital, should treat that gap as a warning about pace: capability demos move fast, dependable daily use moves much slower, and the companies that win are the ones solving the boring reliability problem, not the flashy capability one.
Which of these worries you more: an AI agent with real control over your smart home devices, or one with merged access to your chats and your actual work files?
Sources