AI + Web3 intelligence for builders and creators

Growth Bytes

Issue #003  ·  May 18, 2026  ·  web3matters.xyz

Hermes Beat OpenClaw in 4 of 5 Rounds — Here's What That Actually Means for Your Stack

From Friday's live debate on Web3 Matters

Trav and Q ran a structured, five-round comparison between Hermes and OpenClaw — scored by Claude, double-checked by other tools, and debated live on stream. The results leaned clearly in one direction. But the honest verdict has an asterisk that matters depending on what you're actually building.

The Five-Round Scorecard:

The honest verdict in one line: "Hermes is the better agent. OpenClaw is the more practical one. Whether better beats practical depends on what you're building — this month or next quarter."

Key takeaways from the stream:

Hermes' biggest edge is compounding. Every time you complete a task, Hermes logs it as a skill. It learns your patterns, your preferences, your workflow — and gets faster each session. OpenClaw doesn't do this. Over weeks of use, that gap widens significantly.

The "who am I" file is step zero. Before you give Hermes any tasks, ask your most-used LLM to describe who you are based on everything you've ever discussed. Claude and ChatGPT both do this well. That file becomes Hermes' persistent context — it skips the cold start and starts useful immediately.

Windows setup is painful, but Claude Code handles it. Don't touch the install docs manually. Open a terminal, type claude, and ask Claude Code to install Hermes for you. It handles the WSL/Linux subsystem complexity. One prompt, done.

Security comes before productivity. Q's honest take: running Hermes on your daily machine carries real risk. His rule — keep crypto completely off that machine, define a specific folder Hermes works inside, and make it ask before touching anything outside. Trav is moving his crypto funds off first, then installing next week.

ShipGuard live demo. Trav scanned his public OBS overlay repo on stream and scored 94%. Repo hygiene flagged at 78 (no package.json, no deployment config), but dependencies and secrets both hit 100. The tool is free for public repos — link below.

The Big Idea: The reason builders get obsessed with local agents isn't any single feature — it's the compounding. You master ChatGPT for writing. You add Claude Code for prototyping. Then you layer in Hermes to orchestrate both, learn your workflow, and carry context across everything. Each skill stacks. The system gets better every week you use it. That's the unlock most people aren't talking about clearly enough: it's not about any one tool, it's about building a stack of intelligence that improves itself.

What to do: Ask your primary LLM — right now — to describe who you are based on your conversation history. Save that as a markdown file. That's your Hermes onboarding doc when you're ready to install. Then read Q's security checklist before you run anything locally on a machine that touches real funds.

📺 Full stream: HERE  ·  🎵 TikTok clips: HERE

Elon Flipped on Anthropic, an Unreleased Model Found 27-Year-Old Bugs, and AI Compute Is Moving Off the Planet

The two stories that reshaped the AI power map this week

Musk Leased His Entire AI Data Center to Anthropic — and It's a Direct Hit on Sam Altman

SpaceX — which absorbed XAI and rebranded its AI division — signed a deal to lease the full compute capacity of its Colossus One data center to Anthropic. That's 220,000+ NVIDIA GPUs and over 300 megawatts of capacity, effective immediately. The result for users: Claude Code rate limits doubled across Pro, Max, Team, and Enterprise. Peak-hour throttling removed. Opus API limits raised significantly.

The deal has two layers. First, money: SpaceX is targeting a $1.75–$2 trillion IPO valuation with a roadshow starting June 8, and Anthropic as a named compute customer turns idle GPU capacity into high-margin recurring revenue right before going public. Second, and more importantly — Musk is mid-lawsuit against OpenAI, and Anthropic is OpenAI's biggest direct competitor. As one market researcher put it on X: "Elon's enemy is Sam. Dario's enemy is Sam. Enemy of my enemy is a compute partner." In February, Musk was calling Anthropic "misanthropic" and "evil." This week he said no one set off his evil detector. Three-month total reversal.

Why You Care: If you've been hitting Claude rate limits or getting throttled on Opus, that's fixed. Enjoy the extra capacity — and watch closely, because AI company alliances are shifting faster than product roadmaps right now.

Anthropic's Secret Model Found Thousands of Zero-Days — Including a Bug That Sat Undetected for 27 Years WATCH THIS SPACE

Anthropic launched Project Glasswing: a controlled program giving 12 select organizations — including Apple, Google, JP Morgan Chase, and Microsoft — access to an unreleased frontier model called Claude Mythos Preview. The mission: find and patch critical software vulnerabilities before attackers can exploit them. In weeks of internal testing, Mythos identified thousands of zero-day vulnerabilities across every major operating system and web browser, including a 27-year-old bug in OpenBSD that had never been caught.

The UK's AI Security Institute now estimates that frontier offensive cyber capability is doubling every four months. Both Mythos and GPT-5.5 cleared a 32-step end-to-end cyberattack range in a single month. Anthropic decided Mythos is too dangerous to release publicly — and Dario Amodei publicly called for stronger government regulation, which is rare for a frontier lab CEO to do voluntarily.

Why You Care: The same class of model that's finding your bugs is also being used to exploit them. The offensive/defensive gap is closing at a pace most organizations aren't prepared for — which is exactly why tools like ShipGuard exist as a builder's minimum baseline.

The Agent Framework That Hit 100K Stars Faster Than Any Project in History

Hermes just became impossible to ignore — here's how to actually get started

Hermes

Open-source local AI agent — persistent memory, skill creation, multi-session orchestration

Hermes is a local AI agent framework that runs on your machine and learns from every interaction. Unlike cloud-based assistants, it builds a persistent context of who you are, what you work on, and how you prefer to operate — then creates reusable "skills" from completed tasks so it gets faster over time. It connects to your choice of LLM (Claude, ChatGPT, Gemini) and can orchestrate multiple AI tools in parallel sessions from a single terminal.

This week it hit 100K GitHub stars in seven weeks — faster than LangChain, faster than AutoGPT, faster than any agent framework on record. V0.13.0 just shipped with durable multi-agent coordination, persistent goal tracking, security hardening, and Google Chat support.

Quantic uses it for: orchestrating separate workflows across Web3 Matters, Flare, and ProofRails — each with its own context folder — while running Claude and Hermes in parallel terminals simultaneously.

Trav plans to use it for: content production workflows, newsletter automation, and SEO — once he moves his crypto funds off his daily machine first.

How to get started (the non-painful way):

1. Open your terminal and type claude to launch Claude Code.

2. Ask it: "Can you install Hermes on my machine?" — it handles WSL/Linux setup for you on Windows.

3. While it runs, ask your primary LLM: "Tell me who I am based on everything we've ever discussed." Save the output as a markdown file.

4. Feed that file to Hermes on first launch as your persistent profile. This is what lets it start useful immediately.

5. Set a permissions boundary: tell Hermes it can freely work inside one dedicated folder, but must ask before touching anything outside it.

Free & open-source — 295 contributors, 864 commits in the latest release cycle.

⚠️ Honest caveat: Setup on Windows still requires WSL and is not beginner-friendly without Claude Code's help. And running any local agent with broad file system access is a real security trade-off — do not use it on a machine with crypto funds, private keys, or sensitive credentials until you've defined strict permission boundaries.

Checkout Hermes: HERE

ShipGuard — Quick Follow-Up from Last Week

Free GitHub repo security scanner — now officially named and live

Q's security scanner from last week's emergency episode now has a name: ShipGuard. Trav scanned his public OBS overlay repo live on stream and scored 94% — dependencies and secrets both clean, repo hygiene flagged at 78. The tool is free for public repos. Find it via @0xQuantic on X or the Web3 Matters account. Powered by SHIP.

🔗 Access:  theshipguard.vercel.app

Next 7–14 Days

Next Friday: Neptune Privacy joins the stream. They're building a quantum-safe Layer 1 blockchain combining ZK-Starks and proof of work. With the Claude Mythos findings fresh and offensive AI capability doubling every four months, this is exactly the right conversation to have right now. If you have burning questions about quantum security and blockchain, drop them in the comments or DMs — Trav and Q will ask them live.

SpaceX IPO roadshow kicks off June 8. With Anthropic now a named compute customer, watch how the S-1 frames AI revenue. This filing will be the clearest public window yet into how the SpaceX/XAI/Anthropic compute relationship is valued by the market.

Token-gated content is coming for SHIP holders. Q is building a simple app to unlock research reports, PDFs, and build guides exclusively for SHIP token holders. First drop expected within the next two weeks — including the best practices guide for getting started with Hermes and Claude Code safely.

The question worth sitting with this week: your primary LLM already knows who you are — have you ever actually asked it to tell you?

🔥 ENJOYED THIS ISSUE?

If Growth Bytes is part of your weekly stack, the best thing you can do is share it with one builder or creator who needs it.

🐦 X: HERE
📺 YouTube: HERE
🎵 TikTok: HERE
🌐 web3matters.xyz

We drop every Monday. See you next week.
— Trav & Q, Web3 Matters