A computer the size of a drink coaster, built specifically so your AI agents don't have to share your laptop. That's the pitch behind Lapis One, which Pamir AI launched on September 9, 2026 with a 1:27 demo video that racked up nearly 900,000 views within a day. At $359 for early-bird buyers, it's a real Debian Linux computer — not a phone accessory, not a dev board you have to case yourself — aimed at people who want an agent running continuously, remotely reachable, without tying up the machine they actually work on.
TL;DR
| Question | Answer |
|---|---|
| What is it? | A pocket-sized Debian Linux computer built around running AI agents 24/7 |
| Price | $359 early bird (regular price $599), first 2,000 units |
| Chip | Rockchip RK3576 — 4x Cortex-A72 + 4x Cortex-A53, 6 TOPS NPU |
| Memory / storage | 8GB LPDDR5, 64GB eMMC, expandable via M.2 2230 and microSD (up to 2TB) |
| Size / weight | 85 x 85 x 28mm, 227g — about the footprint of a drink coaster |
| Standout feature | "Agent KVM" — a USB-C port that lets the agent see and control a separate computer's screen, keyboard, and mouse |
| Battery | 3,300mAh, 2-4 hours depending on workload |
| Ships | First batch (140 white units) in October 2026; wider availability in December 2026 |
| Is it just an expensive Raspberry Pi? | No — it's a finished consumer product (enclosure, display, battery, remote access baked in), not a bare board, which is the core of the price debate happening in replies |
What's actually inside the Lapis One
At the center of the device is a Rockchip RK3576 system-on-chip: four Cortex-A72 performance cores paired with four Cortex-A53 efficiency cores, plus an integrated 6 TOPS neural processing unit for accelerating on-device AI workloads without round-tripping to the cloud for every inference. That's paired with 8GB of LPDDR5 memory and 64GB of built-in eMMC storage, with room to grow via an M.2 2230 slot for an SSD and a microSD bay that accepts cards up to 2TB.
The chip choice matters for context: RK3576 is a mid-range ARM SoC already used in a handful of other single-board computers and set-top boxes, not a custom silicon design. Pamir AI's differentiation isn't the processor — it's everything built around it.
That "everything else" is a genuinely complete industrial-design package: an aluminum housing with a plastic top shell, a 240 x 400 pixel Sharp monochrome memory LCD (the same low-power display family used in Panic's Playdate handheld) for always-visible status text without draining the 3,300mAh battery, two MEMS microphones and a 1W speaker with a dedicated XMOS audio DSP, Wi-Fi 6, Bluetooth 5.4, NFC, a fingerprint sensor, a power button, a rocker switch, and a 12-pin GPIO connector supporting PWM, ADC, I2C, UART, and SPI for hardware projects.
The feature actually driving the buzz: Agent KVM
Of the Lapis One's four USB-C ports, one is dedicated to what Pamir AI calls Agent KVM: plug it into a separate computer, and the Lapis One captures that computer's display at up to 2K resolution and 30fps while emulating a full keyboard and mouse over the same cable. In practice, that means an agent running on the Lapis One can watch what's happening on another machine's screen and act on it — clicking, typing, navigating — the same way a human would with physical peripherals, without installing any agent software on the target computer at all.
That's a meaningfully different approach from most current computer-use agent setups, which typically run as software on the same machine they're controlling (a browser extension, an OS-level accessibility hook, or a virtual display). Agent KVM instead treats "controlling a computer" as a hardware-level problem: the target machine just sees a generic keyboard, mouse, and (if needed) a video output — it has no idea an AI agent is on the other end of that USB-C cable. That has real implications for both compatibility (it works with any OS or device that accepts a standard USB keyboard/mouse/display, including ones with no agent-friendly APIs at all) and risk (there's no software-level guardrail on the target machine stopping the agent from doing something destructive — the burden of caution shifts entirely to whatever's running the agent on the Lapis One side).
The other half of the pitch: always-on, always-reachable
Beyond Agent KVM, the Lapis One is built to run continuously and be reached from anywhere. It ships with Tailscale preconfigured, letting you SSH into the device securely from a phone or another computer without manually configuring a VPN or exposing ports to the open internet. Combined with the battery and low-power memory display, the idea is a device you can leave running an agent workload — monitoring something, executing a scheduled task, standing by for a KVM session — and check in on or redirect from anywhere, without your laptop needing to stay awake or your work desktop needing to stay logged in.
Under the hood, Pamir AI built its own OS image for this rather than shipping stock Debian: Lapis OS is Debian 13 (Trixie) on Linux 6.1.141, with a read-only root filesystem verified by dm-verity, full-disk encryption via LUKS2 backed by a TPM 2.0 chip, and signed A/B updates with rollback — the kind of update and integrity architecture more commonly associated with phones or Chromebooks than with hobbyist single-board computers. The company also offers a plain, non-agent-focused Linux image for anyone who wants the hardware as a general-purpose tiny computer rather than an agent runtime specifically.
The $359-vs-Raspberry-Pi debate, playing out in real time
The most pointed reaction to the launch, visible directly in the replies under Pamir AI's announcement, was a direct price comparison: one commenter wrote that "$359 for a Raspberry Pi with a couple of attachments is crazy." Pamir AI's own account replied asking people to "double check our spec before you compare with a rpi" — implicitly conceding the comparison is the obvious first reaction, while pushing back on treating it as apples-to-apples.
Both sides have a point, and it's worth separating them cleanly:
- The Raspberry Pi comparison is fair on raw compute. A Raspberry Pi 5 board costs a fraction of $359, and depending on configuration, its Broadcom SoC can be competitive with or exceed the RK3576's CPU performance for general-purpose Linux workloads. If your goal is purely "cheapest way to get a small Linux box with decent CPU performance," a Pi plus a case is still the better deal.
- The comparison breaks down on what's actually being sold. A bare Raspberry Pi doesn't include a battery, a display, an enclosure, Agent KVM hardware, a fingerprint sensor, or a signed and encrypted OS image maintained by the vendor — you'd have to source, assemble, and maintain all of that yourself, and even then you wouldn't get the USB-C-based KVM capability, which isn't a standard Pi accessory. Pamir AI is selling a finished, opinionated product built around one specific job (running agents continuously and reachably), not a general-purpose development board.
Whether $359 is a good deal comes down to whether that finished-product convenience is worth the premium over assembling the equivalent yourself — which is a legitimate question, not a settled one, and exactly the kind of debate a launch thread with hundreds of thousands of views tends to surface within the first day.
Not everyone can buy one yet
Alongside the pricing debate, a UK-based commenter posted publicly that he was unable to complete a preorder from the UK, referencing that he already owns other niche AI/robotics hardware (a Rabbit R1 and a Reachy Mini) shipped internationally, and asking Pamir AI to fix international checkout "ASAP." As of publication, Pamir AI hadn't posted a public response with an international shipping timeline. This is a common early-hardware-launch pattern — checkout systems, customs/import certification, and initial shipping logistics frequently lag a US-first rollout by weeks or months — but it's a real limitation for anyone outside the US hoping to order on day one.
What this means for builders
- This is a product bet, not a compute bet. If you just want the cheapest small Linux computer, a Raspberry Pi or comparable board remains cheaper on raw specs. The Lapis One is worth its price specifically if you want the finished package — battery, display, KVM hardware, managed OS — without assembling it yourself.
- Agent KVM is the genuinely novel part. Hardware-level computer control (rather than software running on the target machine) is a meaningfully different architecture for computer-use agents, with real compatibility upside and real safety tradeoffs worth thinking through before pointing it at anything sensitive.
- Treat the October/December shipping split as real, not a formality. Only 140 white units are confirmed for the first October batch — if you preorder now expecting a device in your hands quickly, check which batch you're actually in before assuming an October delivery.
- International buyers should hold off and watch for an update, rather than assuming a workaround exists — the UK checkout issue was raised directly to Pamir AI within a day of launch with no resolution timeline confirmed yet.
What to watch next
- Whether Pamir AI resolves international (particularly UK/EU) checkout and confirms a shipping timeline outside the US.
- Independent hands-on reviews once the October batch ships, especially real-world Agent KVM latency and how well the read-only, TPM-backed OS holds up against tampering claims.
- Whether the $359-vs-Raspberry-Pi framing shifts as more people actually use the device, or whether it remains the dominant skeptical take in public reaction.
Related reading
- Perplexity's Portable Computer: RTX GPU Support Confirmed on Linux
- NVIDIA PAIR: Turn Idle Home PCs Into a Personal AI Cluster
- Edge0 Open-Sources a 35B Model That Runs in 2.5GB of Memory
- Analog Devices Buys Alif Semiconductor for $1.35B to Expand Edge AI
- Top 10 AI Tech Gadgets and Directories
This post reflects Pamir AI's public launch announcement, product page, and social media replies as of September 10, 2026. Pricing, shipping timelines, and international availability may change — check pamir.ai directly before ordering.