Talat’s AI meeting notes stay on your machine, not in the cloud

The AI-powered notetaking app Granola, valued astatine $250 million, has go a fashionable instrumentality among tech manufacture founders and VCs. But 1 developer believes there’s request for a much private, local-only alternate that’s disposable for a one-time interest and without a subscription. That’s led to the instauration of a caller Mac app called Talat.

Yorkshire, England-based developer Nick Payne, a self-described machine nerd, says the thought to physique a section AI notetaker came astir mostly due to the fact that of a bid of blessed accidents.

“I deliberation Granola is awesome; it’s a shining illustration of what you tin bash with an Electron app [a model for gathering desktop applications] fixed capable emotion and care,” helium told TechCrunch. “When I archetypal tried it, I was fascinated that it managed to grounds strategy audio connected my Mac without signaling video, which was the modular workaround astatine the time. That led to a ton of research, discovering a comparatively caller and poorly-documented Apple API.”

To marque it easier to enactment with that API (Core Audio Taps, which lets developers pat into a Mac’s audio streams), Payne decided to make an unfastened root audio library, AudioTee.

“During that time, I was dilatory piecing unneurotic a toolkit, but I ne'er recovered thing that felt similar it could basal connected its ain arsenic a merchandise alternatively than conscionable a chill tech demo,” Payne said. “The state-of-the-art hosted transcription models — the aforesaid providers folks similar Granola usage — are incredible, and it’s viscerally chill to spot your code unfurled onscreen successful adjacent real-time. But it ever nagged maine that the tradeoff required providing not conscionable my data, but my audio data; my existent voice,” helium added.

He past stumbled upon a bundle toolkit called FluidAudio, a Swift model that enables afloat local, low-latency audio AI connected Apple devices. It lets you tally small, accelerated transcription models straight connected the Mac’s Neural Engine — Apple’s dedicated hardware for AI processing.

That was the portion that made Payne recognize helium could crook his probe into an existent merchandise — 1 wherever your audio ne'er leaves your Mac, and your transcripts aren’t stored connected different company’s servers.

Talat, which was built alongside Payne’s longtime person and erstwhile colleague, Mike Franklin, is the effect of Payne’s involvement successful the audio space. The effect is simply a 20 MB, one-time acquisition that doesn’t necessitate you to make an relationship oregon adjacent stock analytics information backmost with the developers. There are nary ongoing fees, either.

While immoderate AI notetakers whitethorn person much bells and whistles, Talat offers a streamlined acceptable of features. It captures audio from your computer’s microphone erstwhile you’re successful gathering apps similar Zoom, Teams, Meet, and others, and transcribes it successful existent time. The app tries to delegate speakers successful real-time, but you tin reassign them arsenic needed. You tin besides instrumentality notes, positive edit, delete, oregon divided transcript segments. When the gathering finishes, a section LLM generates a summary with cardinal points, decisions, and enactment items.

The notes, transcripts, and summaries are each searchable successful Talat, too.

In summation to the privateness angle, Payne said the extremity is to springiness users much options.

“We’re leaning into configurability and letting users power wherever their information goes: prime your ain LLM, auto-export to [note-taking app] Obsidian, webhooks that propulsion information retired erstwhile a gathering finishes, an MCP server,” which is simply a standardized mode for AI tools to link to extracurricular information sources, “to propulsion it connected demand,” helium explained.

Under the hood, the AI is simply a substance — “mostly stitched unneurotic and abstracted down FluidAudio,” Payne noted, which helium credits with doing a batch of the dense lifting. For the summarization piece, the app defaults to an Al exemplary called Qwen3-4B-4bit, which tin tally connected adjacent reasonably humble hardware.

However, users tin opt to power that retired to immoderate unreality LLM supplier of their choice, oregon they tin take betwixt 2 Parakeet variants — code designation models developed by Nvidia — oregon constituent it astatine Ollama (a instrumentality for moving AI models locally), giving them much power implicit the experience. In time, Talat volition adhd enactment for much built-in choices, arsenic well, arsenic good arsenic integrations for different apps, similar Google Calendar and Notion.

At launch, users with M-series Mac computers (those moving Apple’s ain processors, starting with the M1) tin download the app and effort it retired for escaped with 10 hours of recordings earlier deciding to purchase.

Talat is available for $49 portion successful this pre-release version, which is inactive nether progressive development.

When the app hits a 1.0 release, the terms volition summation to $99.

Payne and Franklin are bootstrapping Talat and program to support the halfway merchandise a one-time acquisition going forward.