Vercel shipped a public score for whether agents can read your site.
I had already been doing this work by hand on Promptway: llms.txt, feeds, JSON-LD, markdown siblings, an allowlist in robots.txt. Then they put a number on it and a CLI behind it, which is the part I wanted.
The announcement is from @vercel_dev, August 20, 2026.
https://x.com/vercel_dev/status/2090857388924661915
Introducing is-agentic.com, a tool to measure how well agents can read your site. Backed by @oradotai's research, you can run:
Audits with 100+ checks
Visualizations of agents using your site
One-click prompts to fix problems
A CLI for agents
is-agentic.com is the report. Ora is the scanner. That split is load-bearing. Ora runs every check and scores each layer. Vercel groups those results into Essential, Recommended, and Bonus, stores the history, and publishes a shareable page. If you argue with a number, you are arguing with Ora's evidence first and Vercel's grouping second.
The about page is honest about the limits: a score is a public snapshot, not a certification. Automated checks can be wrong. Read the evidence.
The methodology is short. Read it before you treat 74 or 91 as a grade on your company.
Essential checks share an 80-point pool. Recommended checks share 20. Emerging signals can add a bonus, capped at five, and their absence never lowers the score. Partial results get proportional credit. Checks that do not apply are excluded.
That last rule is the one I care about. A publication is not penalized for missing OAuth. A SaaS is not scored as a brochure. API, GraphQL, MCP, and commerce checks only fire when the scan finds that surface. Duplicated check IDs across MCP surfaces are averaged so you do not get double-counted for the same gap.
Start with failed Essential items. Those are the things that stop an ordinary agent: wrong status codes, HTML that is an empty shell, errors that look like pages, content that only exists after JavaScript. Recommended items are for the product you actually ship. Bonus is gravy.
Each report also shows an observed agent journey. One agent, one walk through the site, the places it got stuck. That run is not part of the numeric score. One task is not a sample of every agent. Treat it as a clip, not a study.
Reports cache. A re-scan minutes after a deploy can hand you the old number. Give it about six hours, or wait until the snapshot timestamp moves.
Chris Tate, who works on this at Vercel, posted the joke the same day.
https://x.com/ctatedev/status/2090865331862446088
Best part: is-agentic.com is itself agentic
Your agent can use it to make your site more agentic
That is not branding. The docs describe a stack that looks like the checklist they grade you on.
The CLI is on npm as is-agentic:
npx is-agentic promptway.com
npx is-agentic promptway.com --json
It returns a stored report if one exists, or starts a scan and waits. --json
is the unchanged API body, including structured errors. That is the shape you hand an agent if you want it to close the loop without scraping a HTML report.
Public pages speak markdown. Send Accept: text/markdown
at the same URL that serves HTML and you get Content-Type: text/markdown; charset=utf-8
with Vary: Accept
. If the client rejects every representation they produce, they 406. The HTML report is complete in the first response, so a crawler does not have to run JavaScript to see the score.
The report API is GET https://is-agentic.com/api/v1/report?url=...
. No auth. Errors use RFC 9457 application/problem+json
. It does not start a new scan. 404 means nothing is stored yet; start one from the site. Rate limit is 120 requests per IP per 60 seconds, with IETF RateLimit
headers. OpenAPI lives at is-agentic.com/openapi.json. The RFC 9727 catalog is at /.well-known/api-catalog.
There is a read-only MCP server at https://is-agentic.com/mcp. Three tools: is_agentic_get_report
, is_agentic_get_methodology
, is_agentic_get_developer_docs
. No OAuth, no API key. Registry metadata is at server.json. Hosts that support MCP Apps can render ui://is-agentic/score-report.html
.
It is free. pricing.md is a machine-readable statement, not a pricing page with a toggle. No write APIs, no webhooks, no keys. You run a scan in the browser. Everything else is read.
Run the CLI once. Read Essential failures against a live curl
before you edit anything. Soft 404s and missing Vary
headers are easy to misread from source.
Then hand the report's fix-it prompt to whatever coding agent you use, with one extra instruction: do not invent product surface to chase Recommended or Bonus. If you do not run an MCP server, do not publish a fake one. If you are a publication, document the machine-readable content stack and stop. The applicability model only works if you refuse the checks that are not yours.
I pointed it at this host and wrote the field report separately: is-agentic Scored Promptway 74. Here Is What I Changed. That piece is the 74, the diffs, and the prompt I would reuse. This piece is the product.
I am not reciting Ora's full check list. They own that, and it will move. I am not treating a score as an accessibility audit, a security review, or a ranking guarantee. The about page already said so.
I also have not waited out the cache on Promptway's production rescan. The number I have is the one from before we shipped the 74-post fixes. When the snapshot timestamp moves, that field report is the place to look.
If you only read one page besides this one, make it the methodology. Then run the CLI on your own host.