Disclosure: this article contains affiliate links. If you buy through them we may earn a commission at no additional cost to you. Our recommendations are based on the evidence recorded alongside this article, and prices were verified on the date shown.

What changed and why it matters

A developer posted Clawfight.ai to Show HN, describing it as an experiment in how AI agents behave when they're pitted against each other with human spectators watching (agents.md). The post is credited to wesleyhales and is labeled as the creator's first public writeup of the project — there is no second post, roadmap, or follow-up documentation on record.

That's worth noting for anyone tracking how fast agentic AI side projects move from idea to live demo. The creator says the build ran on Claude Code with permission checks skipped ("dangerously skip permissions"), on a machine built around an Nvidia RTX 5090 GPU. That's a concrete, checkable detail about the tooling and hardware behind the project — not a marketing claim, and not something this piece is estimating.

The detail

For a reader meeting this for the first time: Clawfight.ai is not a packaged product with a signup flow. Per the creator's own description on the agents.md page, it's an experiment exploring how AI agents interact — including rapping against or fighting each other — in front of human spectators. The available evidence does not show a commercial launch, pricing, or user accounts tied to the project, so treat it as a demo, not a service you can buy into.

What is documented is the build process, and it's the more useful part for other builders. The creator says they used Claude Code with permission checks skipped on a machine built with an Nvidia RTX 5090 GPU (agents.md). Skipping those checks means the coding agent could execute file and system operations without per-action confirmation — a tradeoff the creator states outright, not one this piece is inferring.

Because the agents.md page is the only public documentation, several obvious questions have no sourced answer yet: how agents are scored, what models power each side, whether spectators can submit prompts, and whether the project will get a second update. None of that is in evidence, so none of it belongs in a buying decision today.

What it means for you

Watching one developer assemble an agent-vs-agent demo with Claude Code and a single GPU is a reminder that the barrier to shipping an agentic side project is low. The harder, less glamorous problem is what comes after the demo works on your own machine: getting it in front of spectators online, keeping it up, and not rebuilding your deployment setup every time traffic spikes.

If you're already comparing AI coding tools for a project like this, read Best Continue Alternatives: How to Choose Now before you commit to one — it's written for exactly the stage Clawfight.ai's build sits at, choosing the agent tooling before you choose the hosting.

Once the tooling choice is made, hosting is the next decision, and it's a separate one from which AI agent framework you use. Vercel sells directly into this ai-tools space, and since no pricing or plan limit for it is confirmed here, see current pricing at the merchant See current pricing at the merchant before you size a deployment around it.

Best for: builders who already have a working agent demo — local or cloud — and need a fast way to put it online for spectators without standing up their own infrastructure.

Skip it if: you haven't built anything yet. Read the detail section above first, and treat Clawfight.ai as a one-off experiment to learn from, not a template with a documented architecture to copy.

The decision in front of you isn't whether Clawfight.ai is good — there's not enough public documentation to grade it. The decision is whether your own agent project is far enough along to need a hosting answer, and if it is, that's the comparison worth making next.