What is Pocket AI? How the wearable note taker actually works

"What is Pocket AI" gets asked in a lot of shapes. What is Hey Pocket. What does the Pocket AI recorder do. How does Pocket AI work. Is it a note taker, an app, a wearable, or all three. I've had a Pocket on my collar every working day for six months, so this is the explainer I wish had existed the week I was deciding whether to buy one. No verdict here (that's the full Pocket AI review). Just what the thing is, how it works, and where it fits.

What Pocket AI is, in one paragraph

Pocket AI is a wearable AI voice recorder. It's a small device, 52 grams per heypocket.com, that clips to a shirt or hangs on a lanyard. You press a button and it records the conversation you're in. The audio syncs to your phone over Bluetooth, uploads to Pocket's cloud, gets transcribed with speakers labeled, and an AI model writes a summary with action items and decisions pulled out. A few minutes after the conversation ends, the summary is in the Pocket app. It's built for the conversations a laptop can't hear. Phone calls, in-person meetings, a walk with a freelancer, a drive where you talk through a brief out loud.

Here's the spec read, pulled from what heypocket.com publishes and what six months of use confirmed. Storage and battery figures are the vendor's numbers, and they move with firmware, so treat the site as the source of truth.

Who makes Pocket AI, and why it's called Hey Pocket

Pocket is made by Open Vision Engineering Inc., a hardware company in San Francisco. The product is sold direct at heypocket.com, and that domain is why you'll see it written as Hey Pocket, heypocket, or Pocket AI depending on who's typing. Same device, same company, three names. When someone asks me "what is Hey Pocket," the answer is that it's the website, not a different product.

The naming confusion is part of why "is it real" comes up so often in searches. It's a newer brand than Plaud or Otter, the first-generation app was rough (more on that below), and the domain doesn't match the product name. I wrote a separate is Pocket AI legit post that walks through the company, the 30-day return window, the 1-year warranty, and the third-party review record. This post assumes you're past that question and want to know how it works.

How Pocket AI works, step by step

Mechanically it's a four-step loop, and the loop is the same one every AI note taker runs. What Pocket does differently sits above the loop, in steps four and five below.

1. Press the button. The capture button is on the side of the device, single-press and tactile. You can find it without looking, which matters when a call is starting and you have three seconds to decide whether to record it.
2. The device records and syncs. The onboard mic records the audio to the device's storage. It pairs over Bluetooth to your phone and hands the recording to the Pocket app. If your phone isn't nearby, the recording waits on the device until it is.
3. The audio uploads and gets transcribed. The app pushes the audio to Pocket's cloud. A transcription engine turns it into text, labels who said what, and stores the transcript alongside the audio.
4. An AI model you chose writes the summary. Pocket hands the transcript to the model you've selected in settings. It writes a summary with action items, decisions, and structure pulled out. I point mine at Claude Opus. You could point yours at something else, and you can change it later.
5. The capture becomes searchable. The summary and transcript land in the app's history. If you've registered Pocket's MCP server with Claude Code, every capture is also reachable from there, so "find the call where the client mentioned the rebrand" happens in Claude instead of in a vendor app.

The whole loop takes minutes. On a normal phone call the summary is usually waiting by the time I've walked back to my desk. On a long in-person meeting it takes a bit longer, but never long enough that I've had to wait for it.

How Pocket records phone calls, in-person conversations, and offline

The loop above covers the software. The questions I get most are about the capture itself. Does it record phone calls, does it work without a signal, how close does it need to be. Everything in this section comes from the spec list and FAQ on heypocket.com, checked September 18, 2026.

In-person conversations

Two studio microphones handle room audio. Heypocket.com puts the range at up to about 15 meters (49 feet), depending on the environment. Read that as a best case. In my use the device is cleanest on one-on-one conversations and quiet rooms, and a loud restaurant degrades it the way it degrades every recorder in this category.

Phone calls

Pocket has a third microphone for calls, a contact mic. You put the device on the back of your phone and it picks up the call through the phone itself, so the call doesn't have to be on speaker. The device is MagSafe compatible, and phones without MagSafe use the adhesive mount that comes in the box. That's the design answer to a problem software can't solve on a phone, because iOS and Android don't let a note taker app listen to a cellular call.

Offline recording

Pocket records without internet and without your phone nearby. Audio is saved to the 64GB of onboard storage. When the device is back in range of your phone and you open the Pocket app, it syncs and the summaries get built. Transcription and summaries still happen in the cloud, so offline means record now and read later. For a site visit or a flight, that's the right trade.

Knowing it's recording, battery, and charging

Press the side button and Pocket confirms with a light and a vibration. Press it again to stop. The vendor states 4 days of active use on a charge and 1.5 hours to recharge over USB-C. I charge mine overnight and have never run it flat during a working day.

What you get from each recording

Every capture turns into the same set of outputs in the app, and all of it is searchable afterward.

• A transcript with speakers labeled. Pro adds the highest-accuracy transcription and auto-detected speaker names.
• A summary of the main takeaways, written by the model you picked.
• Action items pulled out automatically, which can push to the task apps heypocket.com lists, including Apple Reminders, Google Tasks, Todoist, TickTick, Linear, ClickUp, and Asana.
• A mind map of the conversation's topics and branches.
• Ask Pocket, a chat box for questions across your recordings. The free tier allows two messages a day and Pro is unlimited.

Heypocket.com also lists calendar connections (Google, Outlook, Apple), document destinations (OneDrive, Google Drive, Obsidian), and AI assistant connections for ChatGPT, Claude, and Cursor through MCP. You don't need a subscription for the core outputs. The pricing breakdown lists what works without a subscription and what Pro adds.

The four layers, hardware, app, cloud, and model

It helps to think of Pocket as four separate layers stacked on top of each other, because the reviews you'll read online usually complain about one layer and praise another.

Hardware. The device itself. Mic, button, battery, storage, Bluetooth radio. This is the easiest layer to judge. It's small enough that I stopped noticing it after the first week, the mic is clean on one-on-one calls and quiet rooms, and it degrades in noisy restaurants the same way every device in this category does. I've A/B'd it against Plaud on the same conversation and couldn't pick a winner on raw audio.

App. The Pocket app on your phone (iOS or Android), plus the web, Mac, and Windows apps heypocket.com lists. This is where you read summaries, search history, pick your AI model, and find the MCP setup. The first-generation app was slow enough that I nearly returned the device two months in. Pocket rebuilt it in 2026 and the current version is fast. If you read a review from before the rebuild, you're reading about a different app.

Cloud. Transcription happens on Pocket's servers, not on the device. That's why it needs your phone and a connection to finish the loop, and it's why the storage on the device is a buffer, not the archive. Pocket states data is encrypted in transit and at rest. Those are the vendor's words, so read the privacy policy before you put client calls through it. I did, and I use it on client calls.

AI model. The layer that writes the summary. Most competitors pick this for you and don't tell you what it is. Pocket exposes it as a setting. That one decision is most of why I own this device instead of a Plaud, and it deserves its own section.

Model choice and the MCP server, explained for non-developers

Two features on the spec sheet sound like developer jargon and aren't. Both are worth understanding before you buy, because they're the whole difference between Pocket and everything else in the category.

Model choice means you decide which AI writes your summaries. Heypocket.com describes the product as model-agnostic and lists GPT, Claude, and Gemini options. In practice that's a dropdown in settings. Pick one, and every future capture gets summarized by it. Why care? Because summaries from different models read differently. With Claude Opus, action items get pulled out more precisely and commitments land in their own section, which is the difference between a summary I skim once and a summary I have to re-check against the transcript. And when a better model ships next quarter, I flip the setting and the whole device gets better. With a vendor-locked device you wait for the vendor.

The MCP server is the plainer one once you strip the acronym. MCP (Model Context Protocol) is a standard way for an AI tool like Claude Code to reach into another app's data. Pocket publishes an MCP server, which means Claude Code can search and read your Pocket captures directly. You register it once, with a single command, and from then on your meeting notes are part of the same knowledge Claude can see when you're writing a brief, drafting a follow-up email, or trying to remember what a client said in March. No exporting, no opening the Pocket app. If you don't run Claude Code today, this feature is dormant. If you do, it removes a context switch you'd otherwise pay hundreds of times a week.

How Pocket differs from a voice memo app, and from Otter or Granola

The two comparisons people actually make when they ask what Pocket AI is are "why not just use my phone" and "why not just use Otter."

Versus a voice memo app. Your phone's recorder gives you an audio file. That's the end of its job. Pocket gives you a labeled transcript, a summary, a searchable history, and a path into Claude Code. The other difference is physical and it matters more than it sounds. Recording with a phone means unlocking a screen, opening an app, pressing record, and then leaving the phone face up on the table between you and the person you're talking to. Pocket is a button on your collar. The conversation gets captured because the friction is near zero, and the phone stays in your pocket. The first week I owned it I forgot to bring it twice. By week three the habit had set and I haven't missed a capture I wanted since.

Versus Otter, Granola, and other software note takers. Those tools listen to your laptop. Granola hears the audio your computer is playing during a Zoom or Meet call, and it's excellent at that. Otter, Fathom, and Fireflies work the same laptop-side job. None of them can hear a phone call from the car, a hallway conversation, or a lunch meeting where nobody has a laptop open. Pocket hears whatever's happening around you, calendar event or not. The two cover different halves of the week. I run Granola plus Pocket together, Granola for scheduled video calls and Pocket for everything else. If every conversation you have happens on a laptop, skip the hardware and run Granola. If half your week happens off it, that half is what Pocket is for.

Versus Plaud and other hardware. Plaud Note Pro is the closest thing to a direct competitor. Hardware and transcription are close to parity. Plaud has the more polished consumer app, deeper templates, and stronger translation today. Pocket has model choice on every summary and an MCP server that registers with Claude Code in one command. Plaud shipped its own MCP server in May 2026 and offers model choice on its paid plans, so the gap is narrower than it was, but Pocket's setup is lighter and the model choice isn't paywalled. That's the trade. The long version is at Pocket vs Plaud, and the hardware note taker ranking covers the rest of the field, including TicNote and Limitless.

How the plans work, in shape

You buy the device once. Then there's a free tier and an optional Pro subscription on top, plus an enterprise tier for teams. Heypocket.com states the first 30 days of the subscription are free, starting after delivery, so you get to run the paid features during the same window you'd use to decide whether to keep the device. Free covers light use, and the gating on the paid tier sits on things like how long summaries are kept, how many Ask Pocket messages you get per day, custom templates, and file attachments. Transcription itself isn't the thing being metered.

I'm not printing dollar figures in an explainer because they move with promotions and I'd rather you read the current number than a stale one. The Pocket AI pricing breakdown has the tier-by-tier read and the "should I pay for Pro" math. The one thing I'll say here is the thing most buyers get wrong. The hardware is the entry fee. The subscription is the rent. If you're comparing Pocket to Plaud, compare a year of both, not the two price tags.

Who should buy Pocket AI

Buy it if a meaningful share of your work conversations happen off a laptop. Phone calls, in-person meetings, site visits, drives where you think out loud. That's the job the hardware exists to do, and it does it well.

Buy it without hesitating if you also run Claude Code or care which AI model writes your notes. For that buyer, model choice and the MCP server are the reason the device exists in the shape it does. That's my stack, and it's why I've worn one every working day for six months.

Skip it, or pick Plaud instead, if you want the most polished consumer app and the deepest templates and you don't run an AI stack. Plaud is the safer pick for that buyer and I say so in every comparison I've written. Skip hardware entirely if all your meetings are video calls. Granola will do the job for less friction.

Whichever way you land, the mechanics are the same. Button, record, transcribe, summarize, search. The full review covers what six months of that loop actually felt like.