Pocket Review. The AI Recorder That Earned Daily Carry.

What Pocket Actually Is

Pocket is a small wearable recorder that captures real-world conversation, the half of your week a laptop notetaker can't hear. Press the hardware button, it records through its onboard mic, uploads the audio to Pocket's cloud for transcription, and an AI model writes a structured summary in the app within a few minutes.

It clips to a shirt or hangs on a lanyard and pairs to your phone over Bluetooth. By the time you've put your phone away after a conversation, the summary is sitting in the app with action items, decisions, and structure pulled out. That loop is the same shape every AI notetaker runs.

Two things separate it from the rest of the category. You pick the AI model that writes your summaries instead of accepting a vendor default. And every capture is reachable from Claude Code through a published MCP server, so nothing gets stranded inside the vendor app. Those two decisions are the reason this device is on my bag and not in a drawer.

The Hardware Disappears

The hardware is the easiest part to review because there's little to argue with. Small enough to forget you're wearing it, a tactile capture button you can find without looking, mic quality at parity with Plaud, and a battery that covers a working day with active capture.

The device sits low profile and doesn't catch on jackets. After the first week I stopped noticing it, which is the right outcome. The recorder that calls attention to itself loses to the recorder that disappears.

Mic quality is what the price point should buy. Clean on one-on-one calls and quiet in-person conversation, degraded in loud restaurants and open coffee shops the way every device in this category degrades. I've A/B'd Pocket against Plaud on the same conversation, raw audio and transcripts side by side. Neither was obviously better. The hardware story between them is a tie, and the divergence lives above the audio.

Battery lands where it needs to. A full working day with active capture, standby well past that, and in eight months I've never run out before getting home. I won't quote an hour count because it moves with firmware updates. Check heypocket.com for current numbers.

One honest tax. Hardware only captures the conversations you remember to bring it to. The first week I forgot it on the charger twice and missed two calls I wish I hadn't. By week three the habit set and the problem went away. Plaud has the same tax. Every wearable does.

Model Choice Decided It

Pocket lets me pick the AI model that writes my summaries. I point mine at Claude Opus 4.8. That single setting is why I carry Pocket instead of Plaud, and for anyone running an AI stack it's worth the price of the device on its own.

The summaries read differently than a vendor default would write them. Action items get pulled more precisely. Commitments land in their own section. The difference between a summary I skim once and trust versus one I re-read against the transcript is not theoretical, it's my Tuesday.

The reason it matters compounds. When a better model ships next quarter, I flip a setting and every future capture goes through it. My summaries six months from now aren't bottlenecked by whatever model the vendor picked on day one. On a spec sheet that's one line. In a workflow you run every day for years, it's enormous.

If you don't care which model writes your notes as long as they're good, this section reads like overkill, and that's fine. Plaud's vendor-picked pipeline produces good summaries today. The setting only matters to the buyer who'll flip it.

The MCP Server And Claude Code

Pocket publishes an MCP server. One command registers it with Claude Code, and from then on Claude can search every capture without me opening the Pocket app. The "find the call where the client mentioned the rebrand" query happens inside my working tools, not a vendor silo.

The capture layer is something you touch hundreds of times a week. Behind a vendor app, every touch costs a context switch. Reachable over MCP, the cost drops to zero, and after eight months the saved friction shows up on a calendar level, not just a workflow level.

Worth saying plainly. Plaud shipped its own official MCP server in May 2026, which closed what used to be the loudest gap between the two devices. Integration is no longer Pocket-only. Model choice still is, which is why the section above, not this one, now decides the buy.

The App That Almost Sent It Back

The first-generation Pocket app was rough. Slow search, a capture history that took a beat to load, a couple of summaries that weren't where I expected them. Two months in, I was close to returning the device. Then Pocket rebuilt the app, and the 2026 version is genuinely good.

Capture flow is fast now. Search is quick on the things you actually remember about a meeting. Model selection lives in a settings pane that doesn't pretend the choice doesn't exist, and the MCP documentation ships inside the app instead of hiding on a developer page. It's the app the product deserved at launch.

I'm telling that story because hiding the warts is the worst thing a review can do. Buy Pocket in the first generation and you'd have hit the same friction I did. Buy it today and you won't. A small hardware company that fixes its software fast is worth more to me than one that ships a polished v1 and never updates. Keep it in mind when you read older reviews online, most of them graded the app that no longer exists.

What Other Reviews Say

My take is one long-term test, so check it against the aggregate. On Trustpilot, Pocket holds 4.7 of 5 across 1,600-plus reviews as of July 2026, most of them five stars.

The praise clusters where I'd expect. Summaries, the size of the device, how little the app asks of you. The critical reviews cluster in two places worth taking seriously. Noisy multi-speaker rooms, where transcription can attribute lines to the wrong person or mangle names. And the subscription cost, which some buyers didn't price in before ordering. Both are real. The first is true of the whole category. The second is why the pricing section below tells you to model the subscription into the decision, not just the hardware tag.

Pricing And What You Get

The device is $99 at the current launch price, regularly $199, with free shipping. The free tier is genuinely usable. Pro is $19.99 a month or $199 a year, and new devices come with the first 30 days of the subscription free. Verified against heypocket.com on July 31, 2026.

Model the subscription into the buy. The hardware tag is the small number. If you're going to want the highest-accuracy transcription and permanent summary history, the real cost of year one is the device plus roughly $199, and that's the number to compare against alternatives. This category reprices constantly, so check the source before quoting me.

Pocket vs Plaud, The Short Version

Plaud and Pocket each win the axes their team prioritized. Plaud polished the consumer app, the templates, and translation. Pocket exposed model choice and shipped its MCP server early. Hardware is parity. Pick by which half you care about.

The longer argument, including what Plaud's MCP launch changed and where Plaud genuinely wins, lives in the full Pocket vs Plaud head-to-head.

The Granola Pairing

Pocket covers the off-laptop half of the week. Granola covers the laptop half, listening to Zoom and Google Meet calls with no bot in the room. That's the pairing I actually run, and together they cover the whole week without me thinking about which surface fits which meeting.

If your work is all video calls, skip the hardware entirely and just run Granola. If your week mixes client calls, in-person reviews, and drives, the two-product setup is the honest answer, and each one is cheap enough that the pairing still undercuts one enterprise seat of the meeting-bot suites.

Where It Still Misses

Templates are functional but less refined than Plaud's, with no voice-tag routing. Translation lags for multilingual calls. Transcription runs through Pocket's cloud, so there's no local-only mode. And noisy rooms with cross-talk still produce wrong-speaker attributions.

The template gap is the one I'd bet gets closed. It's the kind of layer a small team ships in a quarter once it's the priority, and when it happens the product is better than Plaud on every axis instead of just the integration ones. Until then, if deep templates and voice tags are your workflow, Plaud is quietly excellent there.

The cloud dependency is a real architectural difference from something like Granola's local capture. My agency calls include numbers that aren't in any deck, and I've made peace with the trade because the summaries need a frontier model anyway. If your data governance can't accept cloud transcription, this category mostly isn't for you yet.