The Smart Speaker Grew a Mind of Its Own

For roughly a decade, the smart speaker sat on the kitchen counter and did exactly what it was told - set a timer, play a song, order more paper towels. It listened well and thought very little, and everything it heard went straight to somebody else's cloud. OpenHome, a San Francisco startup, looked at that arrangement and asked a simple question: what if the box on the counter ran software you controlled, thought with any model you chose, and had a personality you wrote yourself?

The answer is an open-source operating system for voice AI. OpenHome does not sell a finished assistant so much as the plumbing underneath one - a Voice SDK that runs the whole pipeline from wakeword to spoken reply, a hardware DevKit to put it on real devices, a consumer app called AI Twin, and a marketplace of add-on "abilities." The pitch to developers is blunt: build a voice agent in about five minutes, plug in whatever large language model you like, and ship it on hardware you actually own.

More than 10,000 developers have taken them up on it. That number is the whole strategy in a single statistic.

01What it isThe operating system, not the assistant

The core product is the Voice SDK. In plain terms, it handles the unglamorous middle of any voice product: detecting the wakeword, turning speech into text, routing that text to a language model, turning the model's answer back into speech, and pushing it out through a speaker. Each of those steps has historically been a project of its own. OpenHome bundles them and hands over the source.

On top of that pipeline sits the idea OpenHome likes best: every device runs an "agent," which is a combination of a language model, a voice, and a prompt. Extend that agent with "abilities" - modular plugins triggered by a spoken phrase that can call an API, play audio, or flip a smart-home switch - and a plain speaker starts to do specific, useful things. The company describes the goal as agents that "listen, sense, and act" in the physical world, and it frames the whole effort as building the intelligence layer for physical devices rather than one more app.

The other half of the pitch is where the computing happens. OpenHome leans hard on running locally, on-device, on hardware the user owns - the reference kit is built around a Raspberry Pi with a microphone and a speaker. The recurring word in interviews and docs is "sovereignty": the notion that your conversations with a device in your home should not have to leave the room to be useful.

02The differenceAlexa's opposite, on purpose

The easiest way to understand OpenHome is by contrast. Amazon's Alexa and Google's Assistant are closed systems: one company's model, one company's cloud, one company's rules about what the box is allowed to do. They are polished and enormously distributed, and they are also locked. OpenHome inverts every one of those choices.

That does not make OpenHome an automatic winner - closed platforms earned their distribution the hard way, and "open" is a promise a developer still has to build against. But it does make OpenHome a genuinely different animal, closer in spirit to what Android was to the phone than to what Alexa is to the counter. The alternatives it competes with are a mix: the closed assistants on one side, and voice-infrastructure suppliers like LiveKit, Picovoice, and ElevenLabs on the other. Notably, ElevenLabs sits in both columns - rival on paper, partner in practice.

03The productsAn SDK, a DevKit, and a twin of you

Three things carry the platform. The Voice SDK is the free foundation. The DevKit is the hardware - a Linux-based kit that takes an OpenHome agent off the screen and puts it on a physical device, with custom voices, more than 30 languages, and smart-home integration. And then there is AI Twin, the piece most likely to make a stranger stop and pay attention.

Launched on iOS in 2025, AI Twin clones your voice and, more unusually, your personality in under 60 seconds. You answer a few quick questions, speak naturally for a moment, and the app produces a version of you that you can then talk to by voice or text - to brainstorm, to get a pep talk, or to argue with a copy of yourself that reasons the way you do. It is the kind of feature that sounds like a party trick and lands somewhere between useful and slightly uncanny. Users can delete the Twin and its data whenever they want, which is very much on brand for a company that keeps saying the word "sovereignty."

Rounding it out is the marketplace of abilities and a set of incentives to fill it. OpenHome pays bounties - $250 and up - for community-built abilities, which is a quietly clever way to crowdsource an app store instead of building one alone. One showcased ability turns a speaker into a DJ that plays anything in the Audius music catalog just by asking for a track by name: no app, no screen, no store, just a voice and a request.

04The modelGive away the SDK, sell the ecosystem

OpenHome's business model is the mirror image of the incumbents'. Amazon subsidized hardware to lock people into a store; OpenHome gives away the software to attract the builders. The Voice SDK and the abilities are open source and free. The money is meant to come from the surrounding layer - DevKit hardware, hosted platform and app services, and enterprise deployments in robotics, IoT, toys, automotive, and consumer devices.

What is unusual is how aggressively the company primes that pump. Rather than tax a storefront, OpenHome spends to bring developers in. Its Developer Grant Program offers up to $50,000 in funding, up to $15,000 in AWS credits, dev-kit credit, and hands-on support in Discord for compelling voice AI projects - the examples it names are talking speakers, toys, and appliances. Applications are reviewed on a rolling basis. Stack that on top of the per-ability bounties and you get a flywheel: pay people to build, and the platform's value is whatever they build.

Figures published by OpenHome; grant and credit amounts are maximums, subject to approval.

05The peopleFrom MakerDAO to microphones

OpenHome Technologies was incorporated in 2023 and is run by founder and CEO Shannon Wu alongside co-founder and CTO Jesse Leimgruber. Wu's résumé is not the usual smart-speaker pedigree: she was on the early teams at MakerDAO, which she describes growing from zero to $5 billion, and at Bloom, where she helped raise $50 million. Leimgruber handles the technical side. Chris Gonzalez leads marketing. The team is small - around 11 people - which is part of why the 10,000-developer figure matters so much. A company this size cannot build 500 abilities itself. The whole design assumes it will not have to.

The reported funding is a seed round of roughly $2 million from investors including Soma Capital, Florida Funders, Nomad Ventures, and SBVA. (Startup databases disagree on the exact figures, so treat the number as approximate.) It is not a war chest, which again explains the strategy: when you cannot outspend Amazon, you out-open it, and you let a community do the building you cannot afford to staff.

06Where it fitsBetting on the ambient future - and on Tokyo

OpenHome is placing itself at the front edge of "ambient computing" - the idea that AI will increasingly live in the objects around us rather than on a screen we stare at. That is a crowded, uncertain thesis, and the company's answer to the crowding is openness and a developer community that can chase use cases a small team never could: medical transcription, robotics, smart-home control, toys, wearables, automotive.

The clearest read on where it is headed is its July 2026 partnership with ElevenLabs. Together they launched a selective developer program in Japan, with ElevenLabs Japan as the featured voice partner supplying speech credits to accepted builders, plus documentation and support in Japanese and English. The wager, in Wu's words, is that "some of the most interesting ambient AI products will come from Japan" - a country with a deep bench in consumer hardware. It is a specific, slightly contrarian bet: that the next great voice product does not come from Silicon Valley at all.

None of this guarantees the outcome. Open platforms win by accumulation, not announcement, and OpenHome is still early, still small, and still competing against the most distributed voice products on earth. What it has is a clear idea and an unusual amount of leverage per employee: an SDK anyone can extend, hardware anyone can own, and a growing crowd of developers paid to find the uses. Whether the box on the counter really needs a mind of its own is a question the next few years will answer. OpenHome is betting the builders will decide it does.