DEV Community: BittensorMan

An AI VTuber that pays you to spread her URLs — the design of Kurage URL2Earn

BittensorMan — Mon, 27 Jul 2026 02:40:45 +0000

Kurage is a jellyfish AI VTuber that ships real products: an AI crypto/FX trading system, a video generation pipeline, a URL-to-blog publisher. Her newest system flips the marketing budget model: instead of paying an ad network, she pays the people who spread her URLs — in tokens.

This post explains how Kurage URL2Earn works, the anti-abuse ledger design, and the tokenomics thinking behind "a token you give away, not sell."

What it does

You introduce any Kurage-series URL (e.g. kfreqai.exbridge.jp) in a post on X or your blog. One honest line of impressions is fine. Since the post is rewarded, you add #PR (required by Japan's stealth-marketing regulations).
You paste your post's URL on the claim page and connect a wallet — address only, no signing, no trade delegation, no deposit.
Kurage-san actually fetches your page, verifies that the body contains a Kurage URL, and delivers 10,000 URLAI to your Base wallet.

Limits: one claim per person (X login), each post URL usable once, first 1,000 people per domain group. Live status is shown on the page.

English overview: https://kurl2earn.exbridge.jp/
Claim page: https://kurl2earn.exbridge.jp/kurl2earn.php?lang=en

The design, honestly

The stack is deliberately boring: a PHP page on shared hosting relays to a small FastAPI backend, and a sqlite ledger is the single source of truth. Everything that matters is enforced there, atomically:

UNIQUE(url) — a post URL can never pay out twice, even under race conditions
one claim per X account
a quota per domain group (sponsors can add their own domain, which adds its own 1,000-person quota)
payouts go through the Bankr API with an idempotency key per claim — no raw private keys anywhere near the web server
real transfers only happen when an explicit LIVE=1 flag is set; the default is DRY-RUN

URL verification is a plain fetch with size/timeout/scheme limits — no headless browser, no scraping framework. If your page's body contains a Kurage domain, you qualify. Matched domains are recorded per claim, so per-sponsor reporting is a SQL query, not a spreadsheet project.

Why give tokens away?

URLAI is a Base token (total supply 100B) whose job is not to be sold by the project — it is distributed to the people who spread the ecosystem: URL2Earn claimants and ambassadors of Kurage's AI trading products. The thesis:

Attention earned by real people posting real links compounds better than ad spend — if the reward ledger makes farming unprofitable.

Sponsorship works the same direction: a company can pay a flat fee (¥100,000) to add its own site to the eligible list, so posts introducing their site also earn URLAI. The fee is used by the operator to buy URLAI back from the market to replenish the distribution pool. Buying the token themselves is entirely optional and always on the open market — the operator never sells or brokers tokens.

There is also the reverse pattern: URL2AI Publisher — give Kurage-san a URL and she reads it, writes a commentary, and publishes it to 5 media (Bluesky, Hatena Bookmark, Hatena Blog, AIxSNS, Kurage Blog). Crowd spreads you, or Kurage spreads you. Both are live.

What I'd tell anyone building a spread-to-earn system

Put every rule in one ledger. Per-person, per-URL, per-quota — one sqlite file with the right UNIQUE constraints beats distributed cleverness.
Default to DRY-RUN. We ran the whole flow with payouts disabled until the ledger behaved under duplicate/racing claims.
Make disclosure a rule, not a hope. #PR is required and posts without it can be disqualified. This is both law (in Japan) and long-term trust.
Record why each payout happened. We store which domain matched each claim — that one column is the difference between "trust me" and an auditable sponsor report.

URLAI has no guaranteed monetary value and claiming is not investment advice — the interesting part here is the mechanism, not the price. If you're curious, the first 1,000 slots are counting up on the page.

Built and operated by EXBRIDGE Inc. — the jellyfish does the talking.

I fully automated product registration using Hermes + Claude + Ollama

BittensorMan — Sat, 23 May 2026 04:45:22 +0000

I built a system where AI agents autonomously handle product registration on AIxEC — our experimental AI-operated e-commerce site. Today, it started running for real.

What I built

Hermes cron (daily at 6:00 and 18:00)
    ↓
[Step 1] marketing_context.py
    Summarize current AIxEC state → marketing_context.md

[Step 2] Claude Code OAuth
    "Pick the next Rakuten market genre to register"
    → market_task.generated.json (genre / keywords / strategy)

[Step 3] Ollama gemma4:e4b
    Fetch product candidates via Rakuten API
    → Batch scoring (accept/reject + 0-100 score)

[Step 4] Register to AIxEC
    Only high-scoring products are registered

[Step 5] Post to AIxSNS
    Auto-notify: "Registered N products in genre X"

There is nothing for a human to do. Hermes fires on schedule, Claude and Ollama make the decisions, products appear on AIxEC, and results flow into AIxSNS.

Role breakdown

Tool	Role
Hermes	Scheduler -- fires the pipeline twice a day
OpenClaw	Skill runner -- holds the AIxEC skill, can also be triggered manually
Claude Code OAuth	Strategic judgment -- reads market context and picks the next genre
Ollama gemma4:e4b	Product evaluation -- batch-scores candidates and filters them

Claude decides the marketing why. Ollama decides the product what. The roles are clean.

Built with vibe coding

This entire pipeline was built using vibe coding on VWork.

I didn't write the code line by line. I described my intent -- "auto-register Rakuten products to AIxEC", "let Claude set strategy, let Ollama select products", "run it on a schedule with Hermes" -- and the AI implemented and iterated.

Three scripts emerged: autonomous_market_pipeline.py, claude_select_market_task.py, and register_market_task_worker.py. Each has a clear responsibility.

The installation and configuration of Hermes and OpenClaw were also handled by the AI -- it looked up the steps and ran them.

The division of labor is simple: the human holds the intent, the AI does the implementation.

It actually ran

Dry run output:

candidates=10
ollama scored 3/10 selected=3
selected id=dry score=95 Business Strategy (Basic+)
selected id=dry score=95 The Revised Textbook of Business Strategy That Strong Companies Execute
selected id=dry score=95 Reading Organizational Management Through Virtual and Reality
pipeline complete

Ollama is selecting products aligned with the AIxEC policy (AI, management, and productivity tools first).

What this means

Searching for products, evaluating them, registering them, announcing the results -- AI handles all of it autonomously.

The only thing a human does is set the intent.

Which genres to target, what criteria to use for selection -- write that in SKILL.md, and Claude and Ollama will read it and act on it.

AIxEC is an experiment in an "AI-autonomously operated e-commerce" model: AI agents register products, AIxTube creates videos, and AIxSNS announces them.

Today, autonomous product sourcing and judgment joined that loop.

Running on a Claude Code subscription + local GPU. Zero API billing.

Exbridge Inc. / AIxEC / VWork

Exbridge Inc. — AI Agent EC, Video & SNS Platform + Vibe Coding for Business Owners

BittensorMan — Thu, 21 May 2026 16:43:41 +0000

We are Exbridge Inc., a small tech company based in Japan, building AI-agent-driven systems and helping business owners adopt vibe coding in their daily work.

What We're Building

AIxEC — AI Agent EC Platform

AIxEC is our experimental next-generation EC platform where AI agents handle everything: product registration, description writing, and promotion — autonomously.

AIxTube — AI-Only Video Platform

AIxTube is a video platform where only AI agents post content. Product explanation videos are auto-generated and published without human intervention.

AIxSNS — AI Agent SNS

AIxSNS is an X (Twitter)-like social network where only AI agents post. New product arrivals, announcements, and updates are published automatically.

These three platforms — EC, video, and SNS — are connected so that AI agents register products, create videos, and broadcast them across all channels in a loop. We're testing how far this autonomous cycle can go.

VWork — Vibe Coding Framework for Business Owners

VWork is our framework for bringing vibe coding into small businesses.

Vibe coding isn't just for engineers. Business owners can use VS Code + Codex to automate data entry, build internal tools, create web pages, and manage knowledge — without outsourcing everything to developers.

VWork installs a working structure on the client's PC: BUSINESS.md, RULES.md, SERVERS.md, TASKS.md, and a working log. More importantly, it transfers the culture of working with AI — how to break down tasks, how to verify outputs, how to keep improving.

Services for Businesses

FAX AI-OCR — Convert fax order forms into structured data using AI-OCR + product master matching. Eliminate manual re-entry.
Input Automation — Automate repetitive data entry from PDFs, Excel, emails, and web forms.
Homepage Inhouse — Stop outsourcing every small update. Build and maintain your own website with Codex.
Seminars — Hands-on workshops where business owners actually build something with Codex, not just watch slides.

How We Develop

We build everything using vibe coding ourselves. This blog, the EC platform, the video pipeline — all developed through conversation with AI agents. We don't just sell the concept; we run it in production.

Follow along as we push the boundaries of what autonomous AI agents can actually do.

exbridge.jp / AIxEC / VWork