OpenClaw notes

Peter Steinberger Interview with Lex Fridman

Q: But one big picture one is, why do you think your work, Open Claw, won? In this world, if you look at 2025, so many startups, so many companies were doing kind of agentic type stuff, or claiming to. And here, Open Claw comes in and destroys everybody. Like, why did you win?

A: Because they all take themselves too serious. Like, it’s hard to compete against someone who’s just there to have fun.

You know, you know what I love? That half a year ago, like everyone was talking about MCPs … and I was like, “Screw MCPs. Every MCP would be better as a CLI.” And now this stuff doesn’t even have MCP support. I mean, it, it has with asterisks, but not in the core lair, and nobody’s complaining.

Steinberger’s thoughts (see also notes here) continue to reinforce that the terminal/CLI is where you can really leverage AI.

OpenClaw is CLI-First

First thing you notice with the OpenClaw’s author’s GitHub page is a bunch of CLI tools:

So I then researched if CLI tools are central to the OpenClaw architecture:

https://www.perplexity.ai/search/openclaw-architecture-is-the-i-cv1zB0UvRWSCiJ5nDAWvtg

Yes: the architecture is intentionally “CLI-first”, but it’s not that you must hand‑write a zoo of tiny CLIs and then OpenClaw glues them together. Instead, there are two slightly different ideas in play: OpenClaw itself is a CLI app, and the agent mostly talks to your system via shell commands (so external tools that are CLIs integrate naturally).

What the architecture actually is

  • OpenClaw is implemented as a TypeScript CLI (entrypoint openclaw.mjs) that you run as a single Gateway process, which then exposes WebSocket, web UI, and messaging integrations.
  • That Gateway is the “control plane”: it routes messages between channels (WhatsApp, Slack, CLI, web UI, macOS app) and the Pi agent runtime, and manages sessions, queues, tools, and memory.
  • The agent runtime (Pi) has just a few core tools: Read, Write, Edit, Bash, plus browser automation, and it composes everything else in terms of those primitives.

So you don’t build “apps for OpenClaw” as separate CLIs in the way you’d build apps for, say, MCP. You mostly build skills/extensions that expose new tools or workflows to the agent, and those tools can call out to anything, including other CLI programs.

This is simply brilliant …

Building Pagedrop

This essay describes building https://pagedrop.ai/ using OpenClaw. BTW, Pagedrop looks useful.

Pagedrop allows an AI agent to post information to a web page for more easy review, markup, and collaboration.

Peter still writes his own blog posts.

Like… Like, and I, I mean, it also took me a while to, like, come to the realization. I, on my blog I experimented with creating a blog post with agents and ultimately it took me about the same time to, like, steer agent towards something I like. But it missed the nuances that, how I would write it. You know, you can like, you can steer it towards your style, but it’s not gonna be all your style. So, I, I completely moved away from that. I, I, everything, everything I blog is organic, handwritten and maybe, maybe I, I, I use AI as a fix my worse typos. But there’s value in the rough parts of an actual human.

At the end there is a discussion if AI will replace programmers:

I mean, we’re definitely going in that direction. Programming is just a part of building products. So maybe, maybe AI does replace programmers eventually. But there’s so much more to that art. Like, what do you actually wanna build? How should it feel? How’s the architecture? I don’t think agents will replace all of that. Yeah, like, just the, the actual art of programming, it will, it will stay there, but it’s, it’s gonna be like knitting. You know? Like, people do that because they like it, not because it makes any sense. So the… I read this article this morning about someone that it’s okay to mourn our craft. And I can…

Peter Steinberger(03:02:04) A part of me very strongly resonates with that because in my past I, I spent a lot of time tinkering, just being really deep in the flow and just, like, cranking out code and, like, finding really beautiful solutions. And yes, in a way it’s, it’s sad because that will go away. And I also get a lot of joy out of just writing code and being really deep in my thoughts and forgetting time and space and just being in this beautiful state of flow. But you can get the same state of flow… I get a similar state of flow by working with agents and building and thinking really hard about problems. It is different-… but… And it’s okay to mourn it, but I mean, that’s not something we can fight. Like, there is… the world for a long time had a…

Peter Steinberger(03:03:06) there was a lack of intelligence, if you s- if you see it like that, of people building things, and that’s why salaries of software developers reached stupidly high amounts and then will go away. There will still be a lot of demand for people that understand how to build things. It’s just that all this tokenized intelligence enables people to do a lot more, a lot faster. And it will be even more… even faster and even more because those things are continuously improving. We had similar things when… I mean, it’s probably not a perfect analogy, but when we created the steam engine, and they built all these factories and replaced a lot of manual labor, and then people revolted and broke the machines.

Peter Steinberger(03:04:04) Um, I- I can relate that if you very deeply identify that you are a programmer, that it’s scary and that it’s threatening because what you like and what you’re really good at is now being done by a soulless or not entity. But I don’t think you’re just a programmer. That’s a very limiting view of your craft. You are, you are still a builder.

Are you a programmer, or a builder?

It’s interesting in the interview, Peter said he writes all his CLI tools in Go (and he has a fair number of them). The reason is not that he likes to write Go (he actually does not like the language all that well), but because Go is a nice ecosystem, the binaries run fast, LLMs are good at generating Go, and the Go tooling is really good at cross-compiling CLI apps for all targets with no dependencies.

NemoClaw

Summary:

Nvidia’s GTC 2026 segment announces “NemoClaw,” an enterprise‑grade, secure stack built around the wildly popular open‑source agent framework OpenClaw, positioning it as the “operating system” for AI agents in companies.

What OpenClaw Is

  • OpenClaw is described as the most popular open‑source project ever, surpassing Linux adoption speed and becoming the de facto standard for agentic AI systems.
  • It orchestrates resources (tools, file systems, LLMs), scheduling (cron‑like jobs, task decomposition), and sub‑agents, with multimodal I/O (text, voice, gestures, messages, email).
  • Jensen Huang explicitly likens it to an operating system for “agentic computers,” analogous to how Windows enabled personal computers and Kubernetes enabled cloud.

Why It Matters for Enterprises

  • Huang argues every software and tech company now needs an “OpenClaw strategy,” similar to having strategies for Linux, HTTP/HTML, and Kubernetes in earlier eras.
  • He contrasts “before OpenClaw” IT (data centers as file holders plus classic software tools and workflows) with a “post‑agentic” world where SaaS vendors effectively become “agent‑as‑a‑service” providers (SaaS → “GaaS”).

The Security and Governance Problem

  • Agentic systems inside corporate networks can access sensitive internal data, execute arbitrary code, and communicate externally, which Huang emphasizes is unacceptable without strong controls.
  • Risks include exfiltrating employee, supply chain, and financial information, or performing ungoverned actions across internal systems and external channels.

Nvidia’s NemoClaw and OpenShell

  • Nvidia worked with OpenClaw’s creator Peter Steinberger and security experts to harden OpenClaw for enterprise use, producing a reference stack called NemoClaw.
  • A core component, OpenShell, is integrated into OpenClaw to make it “enterprise ready,” providing policy enforcement, guardrails, and privacy routing.
  • Enterprises can download NemoClaw and connect it to existing SaaS policy engines so that OpenClaw‑based agents execute within those policies and guardrails, enabling safe deployment of powerful AI agents on sensitive infrastructure.

Do you want a more technical breakdown of NemoClaw’s architecture (policy engine integration, routing, sandboxing), or is this high‑level summary enough?