Tendrl
The Raspberry Pi is a lovely machine, and it's the reflex answer for anything that sounds like "IoT." But most connected devices don't need a computer. They wake up, read a sensor, send a value, maybe take a command, and go back to sleep. For that, a full Linux box is cost, power, and upkeep you're paying for nothing. People reached for one anyway because the tiny-chip alternative used to feel like a downgrade: you gave up the real database, the easy security, the nice tooling. That's the part that changed.
Head to head
Take the common case: a sensor node that reports readings and takes the occasional command back.
| Raspberry Pi | ESP32 + Tendrl | |
|---|---|---|
| Board cost | ~$15–80 (Zero to Pi 5) | ~$2–6, and ~$1.50 for an ESP32-C3 |
| Idle power | Hundreds of mA, always on | ~10 µA in deep sleep |
| Battery / solar | Wall power, or hours on a pack | Months to years on a cell |
| Cold start | Full Linux boot from an SD card | Wakes, runs, sleeps; no OS to boot |
| Getting it running | Flash an OS onto a memory card, then set it up by hand | Plug into USB, provision from a browser tab |
| Upkeep at fleet scale | OS patches, SD wear, things to crash | No OS to patch, nothing to corrupt |
| The client | Whatever you assemble | Database, offline queue, encryption, file transfer, all built in |
The feature that ends the argument: deep sleep
A Raspberry Pi never really switches off. Even doing nothing, it's a running computer drawing hundreds of milliamps to keep an operating system alive. An ESP32 does the opposite: between readings it drops into deep sleep at about 10 microamps (tens of thousands of times less), then wakes on a timer or a nudge, does its job, and drops back.
That one habit is what lets a device run for months or years on a battery, or indefinitely on a small solar cell, in every place where running a power wire is the expensive part. It's the difference between a gadget you plug in and one you place and forget.
The part that changed: the full client fits
There was always a catch. The good stuff (a real on-device database, secure messaging, an offline queue) took enough memory that you needed a pricier chip with extra RAM to run it. So the honest advice used to be "spend the extra few dollars." Not anymore. The Tendrl client now runs straight from the chip's storage instead of loading into its limited working memory, which frees up enough room that the cheapest ESP32, with no extra RAM, runs the whole thing.
We tested it both ways on the same board, and the result is boring in the best way: same speed, noticeably more breathing room, every feature intact. The practical payoff is a parts bill that drops from about $6 a board to $2 (or $1.50 for the tiny ESP32-C3) across an entire fleet.
AI at both ends: on the chip and in the cloud
Here's the last reason people reach for a Pi: "I want to run AI on it, so I need a real computer." It's the one Tendrl was built to answer, because there are two kinds of AI a device needs and two natural homes for them: machine-learning vision on the device, and reasoning in the cloud. We do both natively.
On the chip: computer vision on OpenMV cameras
This is machine-learning image processing running right on a camera. On an OpenMV camera board (H7, RT1062, AE3, N6), a vision model looks at each frame and answers a question in the blink of an eye, offline, with the image itself never leaving the device: is there a person, whose face is this, did something move, how many items are on the belt. The camera doesn't just record footage; it understands what it's seeing and can act the instant it does. That's what you want when the decision has to be fast, the connection is spotty, or the picture should never be sent anywhere. (This one needs a camera board; a plain ESP32 has nothing to see with.)
In the cloud: Strand
For everything else, the device just sends its data and a Strand workflow does the thinking. The message lands, Claude reasons about it, our Surface scanner checks it on the way in and out, and a command or alert comes back, to the device, a dashboard, Slack, wherever you want it. You build the workflow once, with no AI code on the device, and every board you deploy is connected to it automatically. This works on any board that can send a message, the $2 ESP32 included.
And the two team up to form the loop people buy a Pi to build, minus the Pi: a camera notices a person, sends a heads-up and goes back to sleep, and a Strand workflow decides what to do and answers back. Fast, private decisions on the device; the heavier thinking in the cloud. One platform, from the smallest sensor to the smartest model.
When a Raspberry Pi is the right call
None of this is Pi-bashing; it's about fit. A Pi earns its place when you genuinely need a small computer: heavy AI models, big software, a local server, lots of files on disk. If the thing really is a little server, use a little server. But if it reads, decides, reports, and sleeps, which is most devices, the ESP32 was built for exactly that, and now it gives up nothing to get there. Cheaper hardware, batteries that last years, nothing to patch, and AI at both ends.