Why I Built Pinclaw — An AI Hardware That Does Less
From DJI embedded engineer to AI wearable entrepreneur. A founder's letter on why every AI hardware company makes the same mistake, why AI's real bottleneck is permissions not intelligence, and why Pinclaw chose to do less.
Why I Built Pinclaw
A founder's letter. About an 8-gram piece of hardware, and the big idea behind it.
A Question I Couldn't Shake
By the end of 2024, the AI hardware battlefield was littered with casualties.
Humane had just laid off a quarter of its staff. The Rabbit R1 was collecting dust in drawers around the world. Every product demo I saw followed the same pattern: a beautifully produced concept video, followed by a deeply disappointing reality.
But I couldn't shake a feeling: the direction was right. The execution was completely wrong.
We talk to AI every day — through ChatGPT, Claude, Gemini. But the interface is still a phone screen. Unlock, open the app, type. Thirty seconds of friction every time. Doesn't sound like much, but if you do it 50 times a day, that's 25 minutes spent just getting to AI.
What if you could just talk? Not at your phone. Not at a smart speaker locked in your kitchen. But at something on your body — something that's always there, ready the moment you need it?
That was the question. And what I learned from the wreckage of those who came before me was this: the answer isn't "build a better AI device." It's something else entirely.
What DJI Taught Me
Before Pinclaw, I was an embedded software engineer at DJI, working on autonomous driving systems and AUTOSAR architecture. Before that, I studied microelectronics and integrated circuits at Harbin Institute of Technology.
DJI taught me something most AI companies never learn: hardware is about constraints, not features.
A drone that tries to do everything — longer video transmission range, better camera, obstacle avoidance, longer battery life — ends up doing nothing well. Physics won't allow it. Every gram of sensors is a gram less of battery. Every feature draws power. Every component generates heat. Every capability you add is a new point of failure.
The best DJI products I worked on weren't the ones with the longest spec sheets. They were the ones where someone had the courage to say "we don't need this."
When I looked at the AI wearable market, I saw the exact opposite instinct. Every company was afraid to ship without a screen, a camera, cellular connectivity, and on-device AI. They were building for spec sheets, not reliability.
And reliability is the only thing that matters in hardware you wear every day.
The AI Hardware Graveyard
Let's be honest about what happened.
Humane AI Pin — $699 + $24/month
A projector worn on your chest that could turn your palm into a screen. Revolutionary in theory. In practice: the projector was invisible in sunlight. The battery lasted two hours. It overheated against your chest. The voice assistant was slow. Cellular connectivity was unreliable. Subscription was mandatory. Humane raised $230 million and delivered what The Verge called "the worst product I've ever reviewed."
Rabbit R1 — $199
A handheld device with a so-called "Large Action Model" that claimed to operate apps on your behalf. The LAM turned out to be mostly hot air — most features were just Android WebView wrappers. The device couldn't do most of what it promised. It couldn't even reliably set an alarm.
Friend Necklace — $99
An always-listening necklace that texted you based on what it heard. Creative concept, but privacy concerns arrived immediately. And the core interaction — an AI proactively messaging you — felt more annoying than useful to most people.
Different products, different price points, the same fundamental mistake: they all wanted to be standalone devices.
They wanted to replace your phone. They wanted to pack an entire computing platform into something you wear on your body. And then physics said no.
The Real Bottleneck: Not Intelligence, But Permission
After months of thinking through this problem, I landed on something: the entire AI hardware industry is solving the wrong bottleneck.
The industry is obsessed with making AI smarter. Better models, faster inference, more parameters. And models are getting smarter — at a staggering pace. GPT-4o, Claude Opus, Gemini Ultra — these are genuinely remarkable achievements.
But intelligence without access is useless.
Think about it: what makes an AI assistant truly useful? Not its ability to write poetry or debate philosophy. It's its ability to act in your world — check your calendar, message your team, find the restaurant you went to last month, adjust your thermostat, remind you to take your medication.
An AI Agent that can't touch your data is just a chatbot with a microphone.
And that's precisely what every AI wearable on the market delivers: a chatbot with a microphone. A very expensive, short-battery-life chatbot with a microphone.
The bottleneck isn't intelligence. It's permission.
Apple's Wall
Your personal data lives in two places: your phone and your computer. On the phone side, Apple owns the wall.
iOS locks your contacts, calendar, health data, smart home, photos, and location inside a sandbox. Third-party apps can only access the narrow slice you explicitly grant, and Apple can tighten those permissions at any time. That's good for privacy. But it creates a nearly unsolvable problem for AI Agent products.
This is why Apple can afford to be slow on AI. Apple Intelligence doesn't need to be the smartest AI. It just needs to be the only AI with full access to your iPhone data. No third-party product can compete with it on integration — not Claude, not ChatGPT, not Manus, not any startup.
Every AI app on your phone is trapped in its own sandbox. It can't read your email unless you forward it. It can't manage your calendar unless you use its calendar. It can't control your smart home unless you rebuild your entire setup around its ecosystem.
This isn't a technical limitation — it's a platform limitation, and it's by design.
So I asked: if the phone is a walled garden, where is the open field?
OpenClaw: Another Door
The answer is your computer.
A Mac mini running OpenClaw has something no phone app will ever have: full OS-level access.
It can open Chrome, log into your Gmail, and read your email. It can use VS Code with Claude Code to help you write code. It can operate Figma, Notion, Slack, Excel — anything you can open on your desktop. It can connect to external APIs via MCP. It can run scripts, manage files, and execute workflows.
This isn't "we integrated the Gmail API" — that constrained, permission-gated, narrow wrapper. This is a real computer, logged in as you, with complete access to your entire digital life. Like a trusted assistant sitting at your desk, except this assistant never sleeps and never forgets.
Apple can't control what software you run on your own Mac. Google can't control what your desktop browser can access. Your computer is still yours. And OpenClaw runs on it as an open-source platform that anyone can audit, extend, and customize.
This was the insight that changed everything for me:
An AI Agent doesn't need to live in a wearable. It doesn't need to live in a phone. It doesn't need to live in a cloud sandbox. It needs to live on a real computer with real permissions.
OpenClaw is that computer. OpenClaw's capabilities are Pinclaw's capabilities.
The Last Mile
If OpenClaw is the brain — a full AI Agent running on a real computer with real permissions — what's missing?
You. Specifically, the version of you that's moving through the world. You in the car. You walking the dog. You cooking dinner. You in a meeting where pulling out a laptop would be rude. You need a way to talk to your Agent from anywhere, at any time, without friction.
That's Pinclaw. That's the whole product.
Pinclaw isn't an AI device. It's a communication device. A walkie-talkie that connects you to your AI Agent. You clip it on in the morning, and for the rest of the day, your entire OpenClaw — all the capabilities, all the integrations, all the context — is just one sentence away.
You + Pinclaw → BLE → iPhone → WebSocket → OpenClaw (your computer)
The iPhone in that chain is a relay — it provides network connectivity and runs a companion app for viewing conversation history and settings. But the intelligence, the capabilities, the permissions — all of that lives on your OpenClaw.
This is fundamentally different from anything else out there. Other AI wearables try to pack everything into a small device. Other AI apps try to work within the phone sandbox. We do neither. We connect you to a real computer where AI has real capabilities.
The Art of Subtraction
Once I understood that Pinclaw's job was communication, not computation, every design decision became clear. And almost every decision was to remove something.
| What We Removed | Why | What You Get |
|---|---|---|
| Screen | You have one on your phone, and OpenClaw has one too | Lighter, longer battery, lower cost |
| Camera | Drains battery and raises privacy concerns | All-day battery, zero creepiness |
| Cellular | Your phone already has the best cellular connection | No SIM, no carrier monthly fee |
| On-device AI | AI runs on OpenClaw, not on your chest | No heat, no throttling, unlimited model size |
| Custom OS | Zephyr RTOS is battle-tested and open source | Rock-solid stability, fast boot, minimal attack surface |
What remains is a device that does one thing: carry your voice to the AI Agent and bring the response back to you. Microphone, speaker, button, BLE antenna, battery, clip. That's Pinclaw.
Saint-Exupéry wrote:
"Perfection is achieved not when there is nothing more to add, but when there is nothing left to take away."
Every time someone asks me why Pinclaw doesn't have a screen, I think of that line.
Engineering for Reliability
Subtraction isn't just philosophy — it's engineering discipline. Every component you remove is a component that can't fail.
We chose the Seeed Studio XIAO nRF52840 Sense as our core module. Not because it's the most powerful chip on the market — it's a 64MHz single-core ARM Cortex-M4F, compared to the dual-core 240MHz ESP32 we used in earlier prototypes. We chose it for its integration: PDM microphone, 6-axis IMU, battery charging IC, and BLE 5.4 radio — all within a 21×17.8mm footprint. Every component integrated into the module is one fewer solder joint, one fewer connector, one fewer point of failure.
| Spec | Value |
|---|---|
| Weight | ~8g |
| Dimensions | 28×20×9mm |
| Battery Life (normal use) | 18 hours |
| Wireless | BLE 5.4 |
| Audio Codec | Opus 16kbps |
| Deep Sleep Current | <5μA |
The migration from ESP32 to nRF52840 was a deliberate trade-off: we gave up raw compute (240MHz → 64MHz) and memory (8MB PSRAM → 256KB RAM) in exchange for a 5x reduction in recording power draw (80mA → 15mA) and a 49% reduction in footprint. For a device whose only job is streaming Opus-encoded audio over BLE, 64MHz is more than enough. And the power savings translate directly into the one spec that actually matters in daily use: battery life.
The firmware runs Zephyr RTOS — the same real-time operating system used in industrial IoT, medical devices, and automotive systems. Not because we need that level of certification, but because Zephyr's BLE stack is the most mature and well-maintained in the open-source world. When you press the button on Pinclaw, the audio path from the PDM microphone through the Opus encoder to BLE transmission has to work every single time. No "sorry, please try again."
In hardware, reliability isn't a feature. It is the product.
Two Modes, One Device
Pinclaw ships with two completely independent operating modes, and I consider this one of our most important design decisions.
My OpenClaw — Free
Run your own AI. Deploy OpenClaw on your own Mac, PC, or server. Pinclaw connects to your personal Agent through our cloud relay. You choose the model, you own the data, you pay us nothing. The relay service is free forever.
Pinclaw Agent — $29/month
We handle everything. Don't want to run your own server? We provision a dedicated cloud Agent for you, with access to Claude, GPT-4o, and Gemini. Zero configuration. Clip on and go.
This dual-mode architecture isn't a compromise — it's a conviction. I believe the future of AI is open. People who want to run their own models, control their own data, and build their own integrations should be able to do that without paying us a subscription fee. People who just want something that works out of the box can pay us to handle it.
Same hardware. Same app. The only difference is where the AI runs.
This means Pinclaw isn't locked to any single AI provider. When better models arrive — and they will, every few months — you don't need new hardware. You update your OpenClaw, or we update our hosted service.
What I'm Actually Building
People often ask me: "What is Pinclaw?"
Honestly, Pinclaw is the smallest thing I could build to realize the biggest vision I have.
That vision: everyone should have an AI Agent that genuinely works for them. Not a chatbot that answers trivia. Not a voice assistant that sets alarms and plays music. A real Agent — one that understands your life, can access your tools, and can take action on your behalf. Safely, privately, under your control.
OpenClaw is the platform that makes that Agent possible. The intelligence lives there. The permissions live there. The integrations live there.
Pinclaw is the bridge that makes that Agent reachable — not just from your desk, but from anywhere. Not through a screen, but through the most natural interface humans have ever had: your voice.
I didn't build Pinclaw because the world needs another small gadget. I built it because I believe AI Agents are going to change everything as fundamentally as the smartphone did, and the one thing standing between us and that future is a reliable way to talk to them.
Pinclaw is that reliable way. An 8-gram clip that connects your voice to a fully-permissioned AI Agent. No screen. No camera. No gimmicks. Just a bridge that works, every time, all day long.
OpenClaw is the brain. Pinclaw is the voice.
You say one sentence, and your entire digital world responds.
If you want to try it, Pinclaw hardware is available for pre-order at $129, or download the iOS app first to try it with your phone's microphone — no hardware needed.
Follow me on X / Twitter, or join the Discord community.
The code is on GitHub, MIT licensed, fully open source.