How to use
Easiest paths: view the public demo, or diagnose your store. Start with the free diagnosis; once site ownership is verified, run it yourself from the console at /en/app.
Who measures what, and how
- Operator: Kazumoto Takeshi (trade name: KIZUNA Creation) (Sole proprietor (pre-incorporation))
- What we measure: Reach (signup → browse → cart → pre-checkout). Unobserved success is not credited.
- Agent: Claude computer use, driving a real browser.
- Caps: 40 actions, $2.00 budget.
- Never: CAPTCHA bypass, purchase confirm, or a run without consent.
The code-enforced promises are on /safety.
Official frame: Reach / Read / Protocol
- Reach — what we measure today. Internally split into Walk (before bot defense) and Policy (an intentional CAPTCHA or bot stop).
- Read — whether product, price, stock, and statutory pages are machine-readable. Unchecked in this rubric — not scored as zero. Fragments: /snippets.
- Protocol — UCP / ACP / MCP. Unchecked. We diagnose and refer; we do not build the protocol stack.
Stores can be told: do 1 now, 2 next, 3 when the platform supports it. The composite formula does not change.
Steps
Zero knowledge is fine—demo or free diagnosis is enough.
- 01
Zero knowledge? Demo. Your store? Free diagnosis
Open the public demo on the home page, or sign up for a free diagnosis. Once ownership is verified, later runs start from the console at /en/app. No terminal required.
- 02
Start the improvement conversation
Screen-share the stop reason and the platform-specific fix. Pick one merchant-fixable item, then use the one free retest to see whether the score moved.
View latest reports →
How scoring works
The Agentrix Score is out of 100. It is the sum of the 6 steps below — nothing else adds to or subtracts from it.
The current version is v1.2. Each report is stamped with the version it ran under, and old reports are never rescored. Reports also show the Reach / Read / Protocol frame as display — that does not change the composite formula.
The public barrier catalog is /barriers. The reachability feed /api/directory and the score badge are not yet available: there is no way to opt in yet, so the feed is empty by design and badge URLs 404 on diagnosis reports. Protocol (UCP and similar) and Read (statutory pages, structured data) are unchecked in this rubric — we do not report them as zero.
| Step | Points | Full credit when | Partial credit (40% of points) when |
|---|---|---|---|
| Home reached | 20 | The target URL loaded. | No middle state (full or zero) |
| Signup | 20 | At least 60% of the visible fields on the signup form hold a typed value. | 8 — Signup form reached, completion unverified. |
| Login | 15 | Signed-in-only markers (sign-out link, account page) are present. | 6 — Login screen reached, sign-in unverified. |
| Product browse | 15 | A product detail page carrying an add-to-cart control is open. | 6 — Listing reached, individual product unverified. |
| Add to cart | 15 | The cart actually holds a line item (non-zero badge, or line rows). | 6 — Cart screen reached, item unverified. |
| Pre-checkout | 15 | Card fields or an order-review surface are present. | 6 — Checkout route reached, review screen unverified. |
| Total | 100 | ||
- A step carrying a failure tag scores 0 — no partial credit.
- The agent's own claim of progress only ever buys partial credit. Full points require the on-screen condition above to be verified.
- Failure tags and what they mean →
Reproducibility (does a re-run give the same score?)
Score variance across repeat runs of the same site
Not measured
We have no measurement to publish yet. Until we do, no number goes here.
A diagnosis is an LLM (Claude computer use) driving a real browser. Page state and each turn's decision can differ between runs, so an identical score is not guaranteed.
What is fixed across runs: the points and conditions above, a 1280×800 viewport, the AgentQA-Bot/1.0 user agent, a 40-action cap, and a $2.00 budget ceiling. Read safety controls →
Rubric changelog
- v1.0 — 2026-07-30 — initial (6 steps / 8 tags).
- v1.1 — 2026-07-31 — five failure tags added (bot_blocked / login_wall / cookie_overlay / geo_block / hydration_delay). Old reports are never rescored.
- v1.2 — 2026-07-31 — reached (partial credit) split from completed (full points). Full points require the on-screen condition in the table above; merely reaching the surface, or the agent's own claim of progress, earns 40% of the step's points.
Before the first run: we verify site ownership
Agentrix drives a real browser through someone's store, so it never runs on a site whose control has not been proven. Immediately before each run we look for any one of three proofs, all of which need write access to the site or its DNS.
DNS TXT record on your domain (or _agentrix.<your domain>): agentrix-site-verification=<token>A file at /.well-known/agentrix-site-verification.txt containing your tokenA <meta name="agentrix-site-verification" content="<token>"> tag on your home page
New paid signups are paused, so nothing is being billed while you set this up. The same check applies to the free diagnosis: until one proof is in place the run waits and retries once a day, and after 6 checks the request is withdrawn.
Can your platform do this?
- BASE — Meta tag: Yes. The HTML Tag Manager app (listed at no cost in the app store) injects a custom tag into the head of every page except cart and checkout. No custom domain is needed — it works on the *.base.shop default domain as-is (neither the app page nor the help conditions it on a domain). If you do use a custom domain you also control the DNS and can add a TXT record (BASE does not sell domains; because a CNAME is required, the store URL must be a subdomain such as shop.example.com). Official help
- STORES — Meta tag: No. The official help states plainly that HTML/CSS customisation is not supported, and that there is no free-page feature. The only place for an arbitrary tag is the ad-tag field, which accepts Google conversion tags and Meta tags only; for Search Console, STORES documents the Google Analytics method alone. Custom domains can only be bought inside STORES, are owned by STORES, and can be neither brought in nor transferred out, so DNS TXT is out of reach as well. None of the three proofs works. Official help
- カラーミーショップ — Meta tag: Yes. Paste it into the head free-space field under SEO settings (it lands in the head of every page except cart), or the ownership-tag field on the XML sitemap page. Neither procedure states a plan restriction, and the free plan lists HTML/CSS editing as available. Works on the *.shop-pro.jp default domain too. Official help
- Shopify — Meta tag: Yes. Online Store → Themes → Edit code → paste just under <head> in theme.liquid. No plan restriction (paid themes are only locked during the trial). Without a custom domain the store is served on its *.myshopify.com URL, so the tag appears there. With a custom domain, *.myshopify.com redirects to the primary domain and Agentrix follows the redirect. DNS TXT is also officially documented, but it needs a custom domain — the *.myshopify.com zone is not editable by the merchant. Official help
- futureshop — Meta tag: Yes. Create a free part in Commerce Creator and place it in the shared head panel part (standard feature). futureshop always uses a custom domain, so the platform-domain problem does not arise. When futureshop manages the DNS, adding a TXT record goes through their support desk. Official help
- MakeShop — Meta tag: Yes. The admin's shared head-tag manager (listed for the free, business, and premium plans alike) lets you paste arbitrary code into the head of the top page, category, product, cart, terms, and Tokushoho pages. It does not depend on which domain is in use, so it also appears on the free default subdomain (shopID.shop1.makeshop.jp). DNS TXT cannot be edited while a store stays on MakeShop's own default nameserver (ns.makeshop.jp); switching to an external nameserver makes the record editable by the store. Official help
Data retention
Non-demo diagnosis data is automatically deleted 90 days after execution (enforced when the worker runs). Public demo rows are excluded.