The demo says AI built the league. The evidence stops at the prototype.
A fantasy league is a strange place to settle an argument about software. Wes Roth built one anyway. The video is titled “AI builds EVERYTHING…” The project under it is Blood Bench. The claim he actually publishes is narrower than that title.
I wanted to see whether AI could coach a fantasy football team through an entire league. So I built Blood Bench, using AI to create the code, graphics, and online game, then connected competing AI models to play against each other. Here’s how it came together, what went wrong, and how you can use the same workflow to build your own ideas.
That is an assisted build, a promised account of the failures, and a workflow he says other people can reuse. He also labels the public demo “very much work in progress.” The title is the advertisement. The description is the experiment. A lot of AI-for-business pitches live in the space between those two sentences.
Generation was the easy part to film. Authority is the part that can page you at 2 a.m.
A hosting demo in shoulder pads
Hostinger sponsored the video. In the description, the thanks and the offer sit ahead of the project pitch. The chapter list then hands three consecutive blocks to the sponsor’s stack: Hostinger VPS Setup at 10:32, Deploying With AI at 16:36, and Connecting Hostinger at 19:25, stopping when AI Tournament starts at 22:34. Sponsorship does not make the experiment fake. As structured, the video is a vendor workflow demo wearing shoulder pads.
Chapter titles are an agenda, not a box score. Nothing in the description names the models, the league rules, the season length, the cost, the failures, or a winner. “What went wrong” is promised and not described. “Coach through an entire league” can mean a model setting lineups for months, or a person still making the calls while models drafted the interface. Both readings fit the sentence he published. The description does not show which one happened.
Roth lists a Hostinger API MCP server as a source, beside a chapter called “Connecting Hostinger.” It is fair to infer he wanted a model to reach that vendor’s API, likely for virtual-server or deployment tasks. A link and a chapter title are not a wiring diagram. Whether that server was used, and how, is not established. A demo link does not prove the app is up. It does not prove a model deployed it either.
A protocol is a door, not a project manager
AI automation stops being a drafting assistant the moment a model can act on a system you pay for. That is the decision under the fantasy theme.
The standard here is the Model Context Protocol. Its Python SDK documentation describes MCP as a way to build servers that expose data and functionality to large language model applications “in a secure, standardized way, ” and compares it to “a web API, but designed for LLM interactions.” A server can expose three kinds of things.
- Tools are actions. Create a server, restart a service, send a message. In a company stack, the same kind of slot can write to a CRM, issue a refund, or push a production deploy.
- Resources are data. A config, a log, a customer record.
- Prompts are packaged instructions a host can reuse.
A client connects. The model may use what the server chose to offer. The door is bounded. It falls short of a blank check. It also falls short of evidence that the model designed, tested, and shipped the product on the other side.
“Secure and standardized” is a claim about the protocol. It does not certify a particular virtual server, a particular API key, or an AI agent left running with those keys. An agent that can deploy, change a customer record, or move money is a different risk from a specification. If you are putting this into a business, ask which actions you exposed, who can invoke them, and who can take the keys back.
Roth also links a separate tutorial on installing OpenClaw on a VPS. That does not establish that OpenClaw coached a roster, built this game, or touched Hostinger inside Blood Bench. Paweł, writing at Product Compass, describes OpenClaw as an always-on agent aimed at a user’s wider digital life, and frames the install around not losing data and credentials. An AI agent that does not clock out, on a server you rent, with API access to the rest of the stack, is an access decision. Make it on purpose. A quick demo is not a reason to skip it.
A season is mostly the decision to do nothing
Whether a model can mock up a league site is no longer the hard question. Pages, graphics, a scoreboard that looks alive on a Tuesday: drafting has gotten cheap. Running the league has not.
A fantasy season is mostly quiet authority. Waivers. Trades. Lineup locks. Long weeks in which the right move is to do nothing. That is closer to sales operations than to a landing page: who may discount, when a deal should sit, and which “helpful” change is actually a mistake. An agent trusted with a season would need decision rights that outlast a demo, a memory of the rules, and a human who had already decided which errors are acceptable. If competing models really played, someone still has to referee. Shared rules, a finished schedule, a log of who acted.
A bad graphic is embarrassing. A bad waiver, or a bad discount, moves money. A landing-page demo never has to survive that part. Nothing published with Blood Bench shows that referee or that log.
What is reusable is the discipline, not the theme. Let a model draft. Let a narrow tool, or a person, deploy. Log the calls. Treat the first public URL as a prototype, because the person who shipped this one already called it unfinished.
Questions worth asking before you copy the workflow
-
Did AI build everything?
No. Roth says he used AI to create the code, the graphics, and the online game. That is an assisted build, labeled unfinished, with no split of human work versus model work.
-
Did a model coach a full fantasy season?
Not established. The goal was to test whether AI could coach a team through an entire league. Rules, season length, and any remaining human decisions are undefined, and no completed season is reported.
-
Which models played, and who won?
Unknown. No model names, scores, team counts, or results are in the description. An “AI Tournament” chapter is an agenda item. It is not a result.
-
What is MCP, and did it run Blood Bench?
MCP is a standard way for a model application to reach tools, data, and prompts. The protocol’s Python SDK compares it to a web API built for LLM interactions. Roth links a Hostinger API MCP server, but use is not shown. “Secure” describes the protocol. It does not describe a given server or API key.
-
Was OpenClaw part of this league?
Not established. It appears only as the subject of a separate install-and-VPS tutorial, not as the coach or the builder of Blood Bench.
-
What should a business actually copy?
The split between drafting and authority. Use models to generate a prototype. Do not treat a hosted demo, a sponsor’s deploy flow, or an always-on agent with API keys as proof the job finished. The test is operational: who can act, which credentials were in the room, who can revoke them, and what was logged.
A parallel worth a glance
If the protocol still feels like plumbing, one reader maps the same shape onto an older machine fantasy.