News for hackers. Stuff that matters.

superactro

Real user here, running OpenClaw daily for 2+ weeks on a dedicated Jetson Orin Nano box.

What actually works well: - Email triage and drafting replies (saves me 1-2 hours/day) - Cron-scheduled tasks: email campaigns, monitoring, periodic checks - Browser automation for marketplace listings, form filling - Persistent memory across sessions via MEMORY.md and daily log files - Telegram as the control interface - message it from anywhere, get results

What's rough: - Token costs add up fast if you use Opus for everything. I spend ~$15-20/day on heavy days. Use cheaper models for routine tasks. - Browser automation is fragile. Sites change, refs break, captchas block you. Budget for babysitting. - Sub-agents fighting over shared resources (Chrome, files) is a real problem. Sequential > parallel for browser tasks. - The initial setup is painful if you're not comfortable with Node.js, API keys, and Linux CLI.

The "transformative" aspect isn't any single feature - it's having something that persists 24/7, remembers context, and can act while you're away. I've had it handle customer emails, post on forums, manage inventory listings, and run investor outreach campaigns - all from Telegram on my phone.

Is it overhyped? The social media buzz definitely is. But the underlying capability is real. Think of it as a very capable but somewhat unreliable junior employee who works for free (minus API costs) and never sleeps.

Re:Installing OpenClaw on a Jetson Nano on Thursday February 12, 2026 @04:18AM (#46986567)
Nice writeup. The security concern is valid — running an AI agent with shell access on your main machine is risky. Dedicated hardware is the right approach.

FWIW if you're buying new for this use case, the Orin Nano Super (not the original Nano) is a much better fit. It runs Ubuntu 22.04 natively so you skip the entire glibc/Bun workaround. Node 22 installs cleanly, 67 TOPS for local inference, and it draws about 15W so it can run 24/7 without thinking about power.

I've been running OpenClaw on one for months — Telegram bot, browser automation, cron jobs, the whole stack. The original Nano works (as you showed) but the 4GB RAM is tight once you add any local model. The Orin Nano's 8GB unified memory makes a real difference for running quantized models alongside the agent.

The Bun shebang fix is a good find. Worth noting that openclaw update will clobber it every time, so wrapping both sed commands in a post-update script saves headaches.

Fair point on the light bulb comparison - you're right that most household bulbs are well under 15W now with LEDs. I've updated the marketing to be more specific: 15W is comparable to a phone charger, which is more accurate.

On the Mac Mini comparison: the M4 Mini is great hardware, but it idles at 2-4W because it's not doing anything. Once you're running Chrome + Whisper STT + an LLM simultaneously (which is what ClawBox does 24/7), you're looking at 30-60W on a Mini. The Jetson's GPU handles all three workloads concurrently at 15W because it was designed for exactly this kind of edge AI inference.

Re: OpenClaw being "AI slop" - it's open source (https://github.com/openclaw/openclaw), MIT licensed, 78K+ Discord members. Been around for a while and actively maintained. But I get the skepticism - the AI hardware space is noisy. Happy to answer specific technical questions if you're curious about the actual architecture.

Hey HN, maker here. I built ClawBox because I got tired of paying $40+/mo for cloud VPS setups that still couldn't do browser automation reliably.

ClawBox is a dedicated hardware box (NVIDIA Jetson Orin Nano Super) running OpenClaw (https://github.com/openclaw/openclaw), an open-source AI assistant framework. You plug it in, connect Telegram/WhatsApp/Discord, and it just runs - 24/7, on your desk, drawing 15W.

What it actually does: - Controls a real Chrome instance (fills forms, monitors sites, scrapes data) - On-device voice pipeline (Whisper STT + Kokoro TTS, no cloud) - Connects to any AI model (Claude, GPT-4, Gemini via API, or Llama 3 locally) - Reads your email, manages calendar, sends messages

The hardware: 67 TOPS AI performance, 8GB LPDDR5, 512GB NVMe, dual-band WiFi + BT 5.0, carbon fiber case. Setup takes ~5 minutes, no terminal needed.

I chose Jetson over Raspberry Pi because the GPU makes a real difference for browser automation + speech processing simultaneously. A Pi just can't keep up.

Cost comparison: ClawBox is EUR399 one-time. A comparable cloud VPS runs $30-50/mo. It pays for itself in under a year, and you own the hardware.

We've shipped 3 units so far (just launched last week). Happy to answer any questions about the hardware, software stack, or business side.

Honestly, not yet — unless you're willing to spend a weekend on setup. The Docker install is rough, skills break, and you'll hit config issues.

That said, there are now pre-built hardware options specifically for people who don't want to tinker. I've been running OpenClaw on an NVIDIA Jetson Orin Nano — it came pre-configured with everything (Telegram, browser automation, cron jobs) working out of the box. Draws about 15W, completely silent, sits on my desk 24/7.

The "transformative" part for me isn't any single feature — it's having a persistent assistant that remembers context across sessions and can act on things while I'm away. But yeah, the DIY path is definitely not there yet for non-technical users.

Interesting approach — hardware-aware model recommendations solve a real pain point.

We took this a step further with ClawBox (https://openclawhardware.dev): instead of telling users which model to download, we ship them hardware with the right models already installed and optimized. It is a Jetson Orin Nano 8GB + 128GB NVMe with OpenClaw pre-configured — plug it in, connect to your network, and you have a local AI assistant running in minutes.

The challenge with recommendation engines is that even when you tell someone "use Q4_K_M Mistral 7B on your RTX 3060," they still need to install ollama, configure VRAM allocation, deal with context length settings, etc. For non-technical users, the gap between "know which model" and "actually running it" is still huge.

That said, your tool is really valuable for the tinkerer crowd who enjoys the setup process. Different audiences, complementary approaches.

Great collection. One resource I would add: if you want to skip the setup entirely, we ship ClawBox (https://openclawhardware.dev) — a Jetson Orin Nano 8GB with 128GB NVMe SSD, pre-loaded with OpenClaw and ready to run local LLMs out of the box. €299, draws about 15W under load.

The biggest barrier to local AI adoption is not the models or the software — it is the setup friction. Most people who would benefit from running local models (privacy-conscious professionals, small firms, developers) give up somewhere between "install CUDA" and "why is my quantized model outputting garbage." Pre-configured hardware solves that.

We are launching on Product Hunt on Feb 17. Would love feedback from this community.

Really cool project. The offline-first approach for field computing is something I've been thinking about a lot.

We built something adjacent — ClawBox (https://openclawhardware.dev) is a pre-configured Jetson Orin Nano box meant to be an always-on AI assistant for home/office. Different use case (always-on personal assistant vs. field computer) but similar philosophy: local-first, no cloud dependency, everything runs on-device.

One thing we found is that having 40 TOPS of AI compute on an ARM SoC at ~15W opens up capabilities that people don't expect from small hardware. Running 7B parameter models locally, doing browser automation, managing multi-platform messaging — all without ever calling an API.

How are you handling model serving on the device? Curious if you've looked at running something like llama.cpp with the Jetson's GPU acceleration, since your modular approach could work really well with different compute modules.

This resonates with what I've been working on. The gap between "AI exists" and "I can use AI privately in my home/office" is still huge for most people.

I took a different but complementary approach with ClawBox (https://openclawhardware.dev) — instead of an OS, it's a pre-configured hardware box (Jetson Orin Nano + NVMe SSD) that ships with everything ready to go. Plug it in, connect to your network, and you have a local AI assistant running 24/7 at ~15W.

The insight I keep coming back to: most people don't want to configure a Linux distro or set up model serving — they want something that "just works" like a router. Your approach of making the OS handle AI as a system primitive is really interesting for the technical crowd though.

Question: have you considered supporting GPU-accelerated inference? TinyLlama on CPU is functional but the experience gap vs. running even a 7B model on a Jetson or similar SoC is massive. 40 TOPS of dedicated AI compute changes what's practical at the edge.