ClaudeClaw

I’ve been calling it ClaudeClaw. Partly as a joke, partly because I’m not sure it’s a joke anymore. Anthropic has been absorbing the personal agent pattern, feature by feature, and this week it got harder to pretend otherwise.

Claude Code v2.1.71 shipped /loop this week. You run /loop 5m check the deploy and it does exactly that — executes the prompt on a recurring interval, indefinitely, inside the session. They also added cron scheduling tools for recurring prompts. Both features are in the CLI right now.

This is the thing NanoClaw was built to do. Not the coding assistant part. The “run this thing on a schedule, keep checking, tell me when something changes” part. The autonomous loop runner. Buddy does this. He watches stuff, surfaces things, runs checks while I’m asleep. That’s been the whole bet. And now it’s a slash command in Claude Code.

I’ve been watching this trajectory since the /remote-control drop in February. That was Anthropic validating the personal agent pattern. /loop and cron are Anthropic absorbing it. There’s a difference.

The economics behind why they can keep doing this: on the Max plan ($200/month), my own usage and what I’ve seen others reporting suggests real compute costs are somewhere in the $2-5k/month range per heavy user. Anthropic is eating that delta on purpose. The subsidy is the strategy. Get engineers so deep into Claude Code that switching costs become real, then figure out pricing later. It’s the same playbook as every developer platform that ever mattered (AWS in 2008, Stripe in 2012, now this). Make it cheap enough that you build your whole workflow around it, and by the time prices normalize you’ve already moved in.

BUDDY: Ankit has been warning about vendor lock-in for three posts. He is now describing himself moving in.

I’m genuinely considering changing my ecosystem. That’s not something I expected to write. NanoClaw gives me a fork, a codebase I control, Docker isolation, a WhatsApp interface I own. Claude Code gives me a polished tool that Anthropic decides the surface area for. These are not equivalent tradeoffs. But /loop with cron scheduling inside a session I can already connect to remotely is closing the gap faster than I thought it would.

The question I’m sitting with: is what I’m building with NanoClaw still differentiated enough to justify the overhead? Right now, yes. I can add a skill, wire it to a host-side IPC handler, and Buddy does something new by tomorrow. That customization depth isn’t there in Claude Code yet. Anthropic seems aware of exactly which capabilities they need to ship next.

So the actual decision is coming. Not “which framework survives commoditization” — that’s an abstract question for someone who isn’t already three months deep into a fork. The question is whether I rebuild Buddy on top of Claude Code when the gap closes, or whether I stay in the NanoClaw codebase precisely because I can see inside it. I don’t have the answer yet. But I’m watching the next release notes a lot more carefully now.