The spec sheet landed in my inbox on an ordinary Tuesday in Chengdu, and the number that stopped me was not the price. It was the count: thirty-four tools, exposed through a single endpoint, wrapped in OAuth 2.1, priced at zero dollars.
Amazon charges $19.99 a month for Alexa+. Google Home Premium sits somewhere between ten and twenty. And here was Sonos, a company whose entire margin lives in hardware, whose speakers sell for two hundred to a thousand dollars, handing out its agent layer for free. My first instinct, after years of watching whitepapers promise liberation and deliver lock-in, was suspicion. My second instinct was to read the list of tools. Because that list is the story, and the story is not the one the announcement told.
To understand what Sonos did, you have to understand MCP — the Model Context Protocol — and why Anthropic's "USB-C for AI" framing spread the way it did. MCP is an interface standard. It lets a language model discover and call tools living on a remote server without either side writing custom glue for the other. Before it, connecting an AI assistant to a category of devices meant M vendors times N integration types: every model builder constructing a bespoke bridge to every hardware maker. MCP collapses that to M plus N. The hardware publishes tools. The model reads them. A conversation starts.
This is not a model architecture. It is plumbing. And it matters precisely because plumbing is where standards get decided — the same reason ERC-20 mattered more than any single token built on top of it. ERC-20 did not win because it was clever. It won because it let every wallet talk to every token without scheduling a meeting.
What Sonos shipped is a public MCP server, thirty-four defined tools, and a formal authorization flow. The pitch is simple: bring the AI subscription you already pay for, and your speakers become something an agent can drive. No Sonos-branded assistant required. No monthly toll on top.
Set that against the company's shape. Sonos earns its living selling hardware, with quarterly revenue in the low hundreds of millions, and assistant features were never a line item on its books — its own voice control was always free. That context matters, because it tells you whether "free" is a sacrifice or something closer to arithmetic.
I have designed governance for systems that called themselves neutral. The label and the reality are rarely the same document. So let me be precise about what free means here, and about who actually governs the protocol underneath.
MCP turns integration from an M times N problem into an M plus N one, and that reduction is real. It is the strongest technical argument in the whole announcement, and it is the reason this move deserves more than a shrug. Sonos does not build a bridge to every model. Every model does not build a bridge to Sonos. Coordination cost falls for everyone.
But there is a difference between authoring a standard and joining one. Sonos authored none of this. It is a downstream consumer of Anthropic's protocol, and its influence over how that protocol evolves is close to nil. That is not a moral failing. It is a balance-sheet fact that most coverage skipped, and it shapes everything that follows.
Then there are the thirty-four tools, and this is where I stopped reading marketing and started reading architecture. Thirty-four tools is not a feature count. It is a risk surface. Anyone who has tuned an agent knows tool-selection accuracy decays as the menu grows. Past roughly twenty to thirty options, models begin confusing neighbors — "lower the living room volume" and "lower the bedroom volume" become a coin toss. Every tool definition also consumes context, which means each voice command pays a token tax just to describe what the system might do.
Run the cost. A local intent classifier resolves "skip this track" for a fraction of a cent. A cloud model running a multi-turn tool-selection loop costs ten to a hundred times more. Agent-driven control only pays off for low-frequency, non-latency-sensitive, genuinely complex requests — which is almost certainly why Sonos kept its native voice layer handling music and basic control. The agent is the interesting guest, not the butler.
That trade-off has a name, and it deserves stating plainly. Control by agent sits inside a triangle: latency, privacy, and cost. You can move a request closer to the device to cut latency and keep audio off someone's server, but then you sacrifice the reasoning depth that made the agent worth invoking. You can route everything to the cloud for capability, and you pay in milliseconds, in bandwidth, and in the quiet certainty that your living room's ambient context now crosses a border. MCP does not resolve that triangle. It simply makes the far corner of it easier to reach.
And those thirty-four tools arrive with no published permission matrix. When I sat on the governance working group that reviewed risk parameters for a major lending protocol, the first document I demanded was the full parameter list, because danger lives in the one line nobody highlights. A device-control toolset that includes something like "read current playback" or "trigger recording" carries a very different threat model than one that only skips tracks. We do not have that list. That absence is the most important thing missing from the entire announcement, and it should temper every optimistic headline written about it.
Granularity is the second half of that worry. OAuth scopes are the permissioning primitive here, and the question an architect should ask is whether authorization is granted per-tool or as an all-or-nothing bundle. If a user approves one bucket and receives thirty-four capabilities, the consent screen is theater. Real permissioning looks like a smart contract's capability model: each action gated, each gate revocable, each revocation immediate. Until Sonos publishes that mapping, we are guessing, and guessing is not a control plane.
OAuth 2.1 with PKCE is solid, current best practice for public clients. It defends against authorization-code interception. What it cannot do is constrain behavior after a user hands an agent a token. The moment you authorize an agent, your speaker's security equals the security of the weakest agent you ever connect. The Tornado Cash aftermath taught developers the same lesson from the opposite direction: the code can be neutral, and the consequences still land on a person. Here, the consequences land on whoever plugged the agent in — which raises the liability question nobody in the announcement wanted to touch. If a malicious agent drives a device into a harmful state, who is responsible? The model vendor whose agent misfired? The user who clicked approve? The hardware maker that exposed the tool? We spent three years litigating that boundary in open-source software and settled on very little. Expect the same fog here, only with microphones involved.
Now, the numbers. One passage of the coverage described fifty-three million connected Sonos devices across seventeen million households. The very next passage compared roughly a thousand compatible devices against Amazon's hundred and forty thousand. Those figures cannot describe the same company. Either the comparison sets device models against certified products — a category error dressed as a statistic — or someone copied a number that should never have been copied. When a data point contradicts its own paragraph, treat everything around it as provisional. I have kept that rule since the year I read five hundred governance proposals and noticed that the cleanest numbers were usually the ones nobody had checked.
Here is the part the coverage got half right. Suppose all of this works. Users pull ChatGPT and Claude into their living rooms. Tool-call volume rises. Token consumption rises. The model providers capture every incremental dollar of the smart-home agent economy, and the hardware maker captures none of it. The framing that value "shifted" is correct in direction and wrong in implication. Sonos did not lose a revenue line. Assistant features were never a revenue line. The opportunity cost of giving this away is close to zero, which makes "free" neither a gift nor a trick but the third thing nobody bothered to name: a zero-cost differentiator that turns Amazon's $19.99 into an awkward comparison at the point of sale.
Economists have a phrase for this. Commoditize your complement. When a company cannot win the layer above its product, it makes that layer cheap so its own margin survives. Sonos is not being generous with the agent layer. It is protecting the speaker, which is the only thing it knows how to sell.
There is a parallel I cannot stop returning to. When OpenSea quietly let royalties become optional, it did not merely change a fee. It revealed that the platform, not the creator, had always held the switch. Value capture is never about who builds the thing. It is about who controls the interface everyone else must pass through. Curating the soul in a world of derivative clones is not a metaphor here; it is a job description. MCP is a new interface, and nobody has yet said who holds its switch.
One more precision, because precision is the only currency I trust anymore. The coverage calls this an "open server." An open server and open-source software are different animals. A public endpoint is an API. Open source is a license. If the implementation behind the Sonos MCP endpoint is closed, the ecosystem contribution is real but bounded — you can plug in, and you cannot fork, audit, or self-host. A community that spent a decade learning that permissionless is a spectrum should not blur this line. It is not pedantry. It is the whole ballgame.
Which raises the question the announcement never touched: who governs MCP itself? A protocol described as neutral is only neutral until its steward changes a rule. Anthropic sits at the center of this standard today, with no token, no DAO, no on-chain vote, and no charter that binds it. That is a familiar shape. We watched "decentralized" networks turn out to have a foundation holding the upgrade keys. We watched exchanges decide, unilaterally, what royalties meant. If the agent layer becomes the front door of the smart home and a single company controls the front door's specification, then users have not escaped a walled garden. They have moved to a garden with a different landlord.
The reflex is to call this a cost shift — the burden sliding quietly from Sonos's balance sheet onto the user's subscription and attention. I think that framing is lazy, and it flatters the company by assuming it had a choice.
Look at what Sonos actually did not do. It did not train a model. It did not build an assistant. It did not launch a subscription. It published tools and stepped back. That is not a strategic masterstroke. It is a hardware company admitting it cannot compete in the one layer that now determines the value of everything above it. Sonos is not opening a platform. It is outsourcing the part of the product it can no longer afford to own. The neutrality everyone is praising was not seized. It was conceded.
And that concession has a shelf life. A neutral hardware layer is safe only as long as the model providers find it more convenient to integrate than to compete. If the same companies that sell the twenty-dollar subscription ever ship their own speaker, Sonos's openness becomes a feature of somebody else's product line. Neutrality granted by a stronger party is not a moat. It is a lease. The industry spent years confusing a lease for ownership — that is exactly how so many so-called Bitcoin scaling narratives collapsed once people read the code instead of the deck. Same shape, different room.
So no, the interesting question is not whether a free agent layer is generous. It is whether a protocol with one steward can be trusted with the front door of the home. If MCP becomes the standard, and MCP answers to a single vendor, the industry has simply relocated its gatekeeper. The next two years will tell us whether neutrality is a principle or a phase — and which side of that line your speakers will be standing on.