Hardware

Two boxes. One software stack.

Start lean or launch ready. Both SKUs share the same OS image, API surface and OTA pipeline — so you can mix them in a fleet and upgrade without re-tooling.

AgentBox

Starter

Entry edge AI

50% off launch$798$399 reserve · founding batch

The lowest-friction way to own local AI. Ideal for embeddings, small chat models, workflow routing and offline assistants.

Reserve Starter

✓ RK3588-class SoC · 8-core ARM
✓ 8GB LPDDR · 6 TOPS NPU
✓ 128GB NVMe
✓ 1× Gigabit Ethernet
✓ 2× USB 3.0 · 1× HDMI
✓ Aluminum shell + active cooling
✓ 12V DC · clusterable

Target: 3B-class models, selected 7B quantized

AgentBox · Most popular

Pro

Launch centerpiece

50% off launch$1,598$799 reserve · founding batch

Powerful enough to feel credible, compact enough to sit on a desk. The recommended box for serious single-node inference and small RAG stacks.

Reserve Pro

✓ RK3588 / RK1-class module · 8-core ARM
✓ 16–32GB LPDDR · 6 TOPS NPU
✓ 256GB NVMe + M.2 slot
✓ 2.5GbE networking
✓ 2–4× USB 3.0 · HDMI · Wi-Fi/BT
✓ Upgraded thermal spreader + blower
✓ 12/19V DC · low-latency cluster mesh

Target: 7B quantized models, RAG, agent services

Looking for a specific job? Meet the Developer Edition for local coding models and AgentBox Home for a private family chatbot. Need 32GB+, premium metal and dual-Ethernet for heavier edge mixes? Ask about the Premium build.

Thermal-first design

Designed for the 100th hour, not the first.

Small AI appliances live or die on thermals. AgentBox uses a CNC/extruded aluminum shell as a heat spreader, a quality SoC pad, and active cooling on every SKU — with a vent path that never recirculates hot exhaust.

▪ SoC heat spreader with high-conductivity pad
▪ NVMe thermal pad where the drive sits near the SoC zone
▪ Field-serviceable fan and SSD
▪ Reverse-polarity & surge protection, clean shutdown handling

Full specification

Subsystem	Starter	Pro
SoC	RK3588-class, 8-core ARM	RK3588 / RK1-class, 8-core ARM
NPU	Up to 6 TOPS	Up to 6 TOPS
Memory	8GB LPDDR	16GB or 32GB LPDDR
Storage	128GB NVMe	256GB NVMe + M.2 slot
Networking	1× Gigabit Ethernet	2.5GbE + optional Wi-Fi/BT
I/O	2× USB 3.0, 1× HDMI, UART	2–4× USB 3.0, HDMI, GPIO, UART
Cooling	Aluminum shell + active fan	Thermal spreader + blower
Power	12V DC, protected	12/19V DC, locking, protected
OS	Ubuntu 22.04 LTS (appliance mode)	Ubuntu 22.04 LTS (appliance mode)
Runtime	Ollama · llama.cpp · containerd	Ollama · llama.cpp · containerd
Clustering	Up to 16 nodes	Up to 16 nodes
Mounting	Desktop feet / VESA	Desktop / VESA / DIN-rail

Lock founding-batch pricing

Reservations are fully refundable until your unit ships in Q4 2026.

Reserve yours → Run the ROI numbers

Questions

Hardware questions

Starter (8GB, 128GB NVMe, Gigabit) is the lowest-friction way to own local AI for 3B-class models. Pro (16–32GB, 256GB NVMe, 2.5GbE) is built for 7B quantized models, RAG and agent services on a single node.

The struck-through figure is the full MSRP. To launch the new editions we're selling founding-batch units at or near cost — so the launch price is a genuine discount, not an inflated reference price.

Yes. Both SKUs share the same OS image, API surface and OTA pipeline, and any node can join the cluster mesh — so you can mix them and upgrade without re-tooling.

More questions? See the full FAQ or talk to us.