WeeklyClaw #008 - The Dust Settles. New Setup Guide, a Hosted Alternative, and Anthropic's Most Powerful Model Yet.

It's been four days since Anthropic pulled the plug. The dust is starting to settle.

Some people switched models. Some switched to the API. Some are still figuring it out. I wrote a full guide on exactly what to do - it's live now.

And then Anthropic dropped a model so powerful they won't let the public use it. Because of course they did.

Welcome back everyone, this is your WeeklyClaw.

I published the full guide today. If you've been putting off rebuilding your setup since Friday, this is the one.

It covers exactly how to get running now that the subscription route is dead. Claude API key instead of the Max subscription. Brave Search (free). Groq (free) for lightweight tasks so you're not burning expensive tokens on heartbeats. Homebrew. Install. Connect Telegram. Install Kickstart and QMD.

Under 10 minutes. Every step. No fluff.

Read the full guide on X

If you're setting up for the first time or rebuilding after the ban, start there.

After the Anthropic ban, I've had the same conversation fifty times. People asking "should I set up my own server or use a hosted service?" So let me break it down properly and even give you the best hosting option (in my opinion).

Self-hosting means running OpenClaw on hardware you control. A Mac Mini in your living room. A Raspberry Pi if you're feeling brave. Or just your PC - simple. The software runs on your machine. Your data stays on your machine. You handle updates, config files, API keys, cron jobs, everything.

This is how I run my setup. Mac Mini next to the TV. Always on. Full control over every setting. If something breaks at 3am, that's my problem. But I also know exactly where my data is and nobody can change my billing overnight.

The upside: total control, total privacy, no monthly platform fee. The downside: you need to be comfortable with a terminal, and when things break you're the IT department.

Hosted means someone else runs the infrastructure. You get a web dashboard. You create agents, set up automations, manage everything from a browser. No terminal. No config files. No wondering why your LaunchAgent stopped working at 4am.

The trade-off is obvious. Less control, more convenience. Your data lives on someone else's servers. But you also never think about updates, security patches, or Docker CVEs.

If you're going the hosted route, Claw Headquarters is the one I'd point you to. It's a full multi-agent platform built specifically for OpenClaw. And the founder is a top tier developer - so i’m a big fan of how well it works.

What makes it different from other hosted options:

Here's what I'd actually recommend depending on where you are:

Just getting started? Go hosted. Claw Headquarters gets you running in minutes without any technical setup. Learn what agents can do first. Worry about infrastructure later. This is literally what I did months ago before I got confident enough to buy a Mac Mini and go for it myself.

Already running a setup but tired of maintenance? Migrate to hosted. You'll lose some customisation but gain reliability. No more 3am debug sessions because a config file got corrupted. Easiest way for busy folk to keep going with OpenClaw.

Running multiple agents for business? This is where hosted really wins. Managing 3-5 agents on self-hosted hardware means juggling multiple processes, memory limits, context windows, and crash recovery. A hosted platform handles all of that.

Power user who wants full control? Stay self-hosted. Read my setup guide above. You'll save money long-term and have complete ownership of everything.

The Anthropic ban forced everyone to rethink their setup anyway. If you're rebuilding, it's worth asking whether you want to rebuild on your own hardware or let someone else handle that part.

Check out Claw Headquarters

The model Anthropic accidentally leaked two weeks ago is now officially out. Sort of.

Claude Mythos Preview launched yesterday as part of Project Glasswing - a cybersecurity initiative. It's only available to about 40 companies including Apple, Amazon, and Microsoft. Not you. Not me. Not anyone running OpenClaw.

Why the limited rollout? Because Mythos is apparently so good at finding and exploiting software vulnerabilities that Anthropic is worried about what happens if hackers get access. CNN's headline: "could let hackers carry out attacks faster than ever." The NYT called it a "cybersecurity reckoning."

Steinberger's quote still rings: "First they copy some popular features into their closed harness, then they lock out open source."

Mythos will matter eventually. When it goes public it'll change what agents can do. For now it's a headline and something to talk about on X - maybe even just Anthropic being OTT to recover their media damage from the OpenClaw bans.

TechCrunch · CNBC · NYT · Axios

The ban happened Friday. By Monday, every OpenClaw community was full of "what now" threads. Here's what's happening.

The Claude API route is what most serious users are doing. Same Claude models, pay-per-token instead of subscription. One user said they accepted Anthropic's $100 credit and burned $20 in 24 hours just on heartbeats. If you go this route, watch your spend and route simple tasks to a cheaper model. I’m actually sticking to this at the moment too.

GPT-5.4 and Codex are the biggest winners from the ban. OpenAI now has native computer-use capabilities and with Steinberger at OpenAI, they're positioning hard as the "harness-friendly" alternative. If you're switching away from Claude entirely, this is probably where you land.

Gemini 3 Flash and Pro are getting more attention than expected. Massive context windows. Competitive pricing - Flash at $0.50/$3.00 per million tokens. Good for research-heavy and analysis agents. Quietly improving.

Kimi K2.5 is number one on the PricePerToken community vote leaderboard - 576 votes, well ahead of everything else. At $0.38/$1.90 per million tokens it's incredibly cheap. The community rates it highly for daily agent work but it won't match Opus for complex reasoning. For routine tasks though, hard to beat on value.

Local models are getting real attention. Qwen3.5 27B on a Mac Mini. Zero API costs. The hybrid setup - expensive model for thinking, local for routine - is the move if you want to keep costs down.

Chat-native /tasks command. Background task board in your conversation. AWS Bedrock Guardrails for enterprise. Smarter agent failover.

Bundled provider support. New plugin onboarding flows. Stronger exec approvals. Security fixes.

Built-in video and music generation. Multilingual Control UI. Structured task progress. Reddit (48 upvotes): "The new OpenClaw version looks really solid."

Three releases in four days. The project didn't slow down because one provider changed their billing.

Thanks for reading WeeklyClaw #008.

Shorter one this week. Two paths forward: self-host with the guide, or go hosted with Claw HQ. Either way your agents keep running and if you’re going for self hosted, i wrote a banger article just for you.

Same time next week. 🦞