Product1 distinct publisher3 min readPublished
The London startup, founded by two former Nothing staffers, runs its transcription on Google's Gemini and wants screen-snapshot permission now, on the bet that the real blocker is the coworker within earshot.
The Product Desk · Product desk

Compiled by The Product DeskSomething wrong?How this is made
The reason a dedicated microphone is a real product idea and not just a hardware founder's reflex is signal-to-noise. At the volume where the person beside you cannot make out your words, a laptop mic an arm's length away is mostly recording the room. That is the wall Zhu describes hitting at Nothing: too quiet for the Mac to catch, too loud for the desk next to him [5]. His workaround, holding an earbud in front of his mouth to demonstrate on a video call with Wired, is also a sketch of the competing form factor [5]. The device already in a buyer's ear only needs a whisper mode for the Relay Q's argument to get harder.
What a closer microphone does not fix is everything downstream of the waveform. Relay's engine is Google's Gemini [10], and Wired's reviewer rated its transcription below Google's own Rambler on the Pixel 11, with punctuation it sometimes gets wrong [16]. Of the four things Wired describes asking Relay to do, exactly one is reported without a reliability caveat: the Google Calendar event worked, the Slack skill gave trouble, the Maps link arrived but is called hit or miss, and an emoji request produced the literal words "heart emoji" [23][12][19][18]. No microphone routes a Slack message to the right recipient.
So the thing being asked of a buyer today is not hardware money. It is a system hotkey, Fn by default, in every text field [9], microphone access, and, if the Skills are the point, permission for the app to snapshot the screen each time it activates [13]. Relay says the data sits locally and only transits its own servers on the way to Google's, with nothing retained [14]. Wired still files the arrangement next to Microsoft's Recall on trade-offs [15]. The security review happens in this budget year; the Relay Q arrives in early 2027 [2], on one of the four platforms Relay names, with Windows, iOS and Android still listed as plans [22][6].
For anyone weighing a pilot, the tally that settles it costs a week. Every dictation attempt somebody abandoned, sorted into two piles: gave up because of who was in earshot, or gave up because the output needed more cleanup than typing would have. A team whose first pile is fat has the problem Relay is building for, and its only real question is whether a 2027 delivery date is worth a 2026 permissions grant. A team whose second pile is fat has a model-quality problem, and a microphone at the lips does not touch it. Teams with both piles full should notice which one has a fix that ships now, and it is the software, running on someone else's model, with the accuracy Wired measured [16].
Ranked by verification strength, evidence, and original report placement.
Wired had trouble with Relay's Slack skill, but no trouble getting it to craft a calendar event in Google Calendar.
Relay is debuting its software today, while the hardware it calls the Relay Q will not be available until early 2027.
Relay inserted emoji inconsistently in Wired's tests: asked for a heart emoji it did it flawlessly, but when the reviewer tried to tack that request on at the end of a sentence it wrote out the words "heart emoji".
Wired sent a message via Relay about a restaurant and in the same breath asked for a Google Maps link to the spot, which Relay provided, though the reviewer says this can be a little hit or miss.
During setup the user chooses a hotkey that triggers Relay in any text field, with Fn as the default; pressing and holding it with the cursor in a text field starts dictation.
Relay's voice-to-text engine is powered by Google's Gemini models.
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One publisher's beta hands-on, no independent measurement
The record is a single Wired first-person test of a pre-release macOS build plus interview quotes from one co-founder. The reporting is specific and includes named failures, which raises its usefulness, but there is no second publisher, no quantified accuracy or latency figures behind the Pixel Rambler comparison, and no verification of the vendor's data-handling statements or the 2027 hardware date.
Day-one availability on one platform, no usage evidence
The only adoption signal is availability itself: the software launches on macOS on the day of publication, three named platforms remain plans, and the hardware that carries the product thesis is roughly seventeen months out. No user counts, customer names, download figures, pricing or deployment disclosures appear in the supplied source, so measured adoption is essentially limited to the launch event.
Keyboard-free framing runs ahead of a beta app and unshipped hardware
The framing — stop touching your keyboard, whisper anywhere, how long can you go without a keyboard — sits well ahead of what is demonstrated. What exists today is a macOS-only beta whose transcription the reviewing publisher rates below Google's own phone dictation, whose agentic Skills failed or wobbled in three of four exercised cases, and whose central premise (a whisper microphone that solves the office volume problem) is not purchasable until early 2027. The overstatement is moderate rather than severe because the same publisher documents the caveats rather than hiding them.
Access-driven pre-announcement with vendor-supplied assurances
Multiple incentive pressures are visible in the record itself. Relay gains from a launch-day story and from pre-announcing 2027 hardware to establish a category position against Wispr Flow and Google before it can ship. The publisher is working from privileged pre-release access and a founder interview, which shapes what is testable. The privacy-critical facts — local storage, non-retention, transit-only handling — are attributed to the company with no independent check, and the engine is rented from Google, giving Relay an interest in framing the dependency lightly. Offsetting this, the reviewer publishes unflattering comparisons and named failures.
Firsthand detail, single publisher, pre-release product
Confidence is moderate. The observable facts — macOS-only launch, Fn hotkey, Gemini engine, Skills, microphone plus screen-snapshot permissions, and the specific test outcomes — are reported firsthand and internally consistent, so the descriptive layer is reasonably solid. Everything forward-looking or trust-dependent is weaker: one publisher, a beta build, an unverified data path and a 2027 hardware date with no corroboration.
product
Proton's Yen stops refusing AI and ships Lumo, moving the privacy fight to terms1 distinct publisher
product
Four leaderboards, four denominators: what you buy when you standardize on a coding agent1 distinct publisher
product
A judge is about to price 34 years of Spirit Airlines paperwork at under two cents a record1 distinct publisher
product
The dirtiest part of the AI gas buildout is a turbine spec, not a nameplate1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 27, 2026