What is OpenClaw?

OpenClaw is the personal AI operator I run on a small machine in my home office. It is not a product. There is no waitlist, no SaaS, no OpenClaw.ai. I built it for myself and I am writing this so I have something to point at the next time someone asks.

What it actually is

A small box at home running a few long-lived processes:

Around fifteen scheduled jobs that wake up on a timer.
Open-source models running locally on a Strix Halo workstation — Qwen 122B for text and Qwen-VL for images.
Free-tier inference from NVIDIA NIM (GLM 4.7, Kimi K2) for the jobs that need a hosted model.
A small team of specialised agents that hand work to each other — one for research, one for coding, one for review, one for orchestration.

The jobs do narrow, repetitive things: drafting morning email replies, watching a few stock positions for technical entry zones, posting market notes when prices move, running a weekly goal review, flagging when one of the jobs has silently broken.

What it costs to run

Effectively zero. Almost everything runs on free NVIDIA NIM, on local inference, or on a small VPS I already had. The jobs themselves are shell scripts, cron entries, and code I have been rewriting.

What it has taught me

Free models do most of the work fine. I had been over-specifying every job with frontier models. Swapping the bulk of them to free NIM tier and local Qwen took the bill from small-but-real to nothing, with no quality drop on the kind of tasks I was running.

Specialised agents beat one general agent. One agent doing nine kinds of job loses context fast. Splitting into a researcher, a reviewer, a coder, and an orchestrator handing work to each other works better. Team design matters more than model choice.

Most operational work is observability, not AI. The hardest failures are silent ones — a job that quietly stopped because an environment variable was not being passed through to a subprocess, a token that expired in the middle of the night, an agent that confidently produced a citation that did not exist. None of these throw errors. All of them need monitoring built from outside the agent. Most of my engineering time goes into noticing when work is not happening, not into the agents themselves.

How this connects to Altronis

When a client asks me to design an agentic AI system for their internal operations, I am not reasoning from theory. I am reasoning from a small version of one I have been running on my own time. The systems Altronis builds for clients are larger and more deeply integrated, but the principles are the same: schedule what you can, observe everything from outside, give each agent a narrow role, and do not pay for frontier reasoning when a small model will do.

If you are figuring out where agentic AI fits in your operations — and where it does not — that is the conversation I am interested in having.

Frequently asked

What is OpenClaw and how does it differ from Claude or other agents?

OpenClaw is an autonomous AI operator — runs continuously on your machine, executes scheduled tasks, manages a memory palace, and orchestrates other AI tools. Claude is a single-conversation assistant; OpenClaw is the loop that wraps Claude, Goose, local LLMs, and ~50 skills into a 24/7 worker.

Can OpenClaw be deployed at an enterprise?

The current build is single-operator (one Zach, one machine). Enterprise deployment is in the roadmap and would require multi-tenant memory isolation, role-based skill access, and audit-log export. Reach out via altronis.sg if you want to be part of the early enterprise design.

What is the relationship between OpenClaw and Lyra?

OpenClaw is the operator layer (the brain that decides what to do). Lyra is the application layer (the CRM+ERP the operator drives). They are designed to compose: OpenClaw decides 'invoice this customer'; Lyra is the system the operator clicks.