The Ora scanner bot

OraBot/1.0 (+https://ora.ai/bot)

What it is

Ora measures how ready a website is for AI agents. When someone requests a score for your domain, our scanner fetches your public pages and the artifacts agents rely on: robots.txt, llms.txt, OpenAPI specs, MCP manifests, structured data. It visits on demand rather than on a schedule, and a scan is a bounded burst of requests with hard timeouts and body-size caps.

How to verify it

A User-Agent string proves nothing, since anyone can copy it. So our scanner signs its requests with HTTP Message Signatures (RFC 9421, the Web Bot Auth profile), using an Ed25519 key. Each request carries Signature, Signature-Input and Signature-Agent headers.

Every request that carries our User-Agent is signed, so you can verify it instead of trusting the string. Treat unsigned traffic on our User-Agent with the same suspicion you would treat any other unverified bot.

The keys themselves live at https://ora.ai/.well-known/http-message-signatures-directory. Anything claiming to be us that fails verification against those keys is not us, and you should treat it like any other unidentified bot.

The requests we do not sign

One part of a scan checks whether your site is reachable to AI crawlers at all, so during a scan we send a handful of requests carrying other crawlers' User-Agent strings (GPTBot, ClaudeBot, PerplexityBot and similar) at your homepage, and we record how your site answers.

Those requests stay unsigned. Signing them as Ora would contradict the User-Agent they carry and destroy the measurement, which is the whole point of sending them. We would rather state this plainly than have you discover crawler traffic from our range and read it as evasion. It is a small number of requests per scan, only ever against the homepage.

Personal Agent Protocol checks

The Personal Agent Protocol (PAP) lets a company publish a file at /.well-known/poppy.json so personal agents can find its sign in and conversation endpoints. When a scan finds that file, OraBot checks it against 1.0. Ora follows version 1.0, the latest published version (2026-11-14). Ora runs the scanner, and you can reach the people who run it at hello@ora.ai.

Tier 0: plain reads of public files

The first request goes to every domain we scan. The others happen only when the file lists something to fetch, and each line says when. None of them carries a credential.

  • GET https://{domain}/.well-known/poppy.json, on every scanned domain. (T0-01)

  • GET https://{domain}/.well-known/oauth-protected-resource, when the file is served by the host its organization.domain names. (T0-07)

  • GET https://{www or apex sibling}/.well-known/poppy.json, when the requested host has no file and its homepage resolved to the sibling. (T0-SIB)

  • GET each supported_versions URL in poppy.json, when the file lists one on the host that served it; up to 3, no redirect followed. (T0-27)

  • GET the issuer's RFC 8414 metadata URL, when the domain's protected resource metadata names the issuer (T0-07). (T0-09)

  • GET each other domain the issuer binds (its protected_resources): /.well-known/poppy.json, and /.well-known/oauth-protected-resource when that file names no auth.issuer, when the issuer binding holds and the issuer lists other domains, at most 3. (T0-23, X-04)

  • GET each listed OpenAPI document, when the file lists OpenAPI documents, at most 3. (T0-17)

  • GET each MCP server's RFC 9728 metadata, when the file lists MCP servers, at most 3, each with at most two GETs. (T0-18)

  • GET each ucp entry's endpoint, its UCP profile, when the file lists a ucp agent entry on the host that served it; up to 2, no redirect followed. (T0-40)

Tier 1: junk requests to the endpoints the file lists

Tier 1 probes run only when the request that starts the scan asks for them. Ora's scan page and report rescans on ora.ai ask for them, as do a check rerun in Ora Studio (only on the organization's own domain) and a debugging tool Ora's engineers run by hand. Any caller of Ora's scan API can ask for them too. Scans that do not ask never send them, and the MCP tools, scheduled jobs and bulk scans never ask. Each request below carries no valid credential, and none can create anything on your side: a correct site refuses each one, except the browser join page, which it serves to any visitor and Ora only reads. A probe follows a redirect only once, and only a 307 or 308 that stays on the same registrable domain. That second request counts toward the request cap.

  • POST to the token endpoint in the issuer's metadata, when the Company has an issuer whose metadata has a token endpoint on the scanned domain; a token endpoint on another domain is not probed. It carries a refresh_token grant with a refresh token nobody issued and a client assertion nobody signed, with client_id https://ora.ai/poppy/probe-client.json, never re-sent after a redirect. A correct site answers invalid_grant, invalid_client or unauthorized_client, or a DPoP proof error; a 403 extension_required reads not tested. (T1-01)

  • GET to {endpoint}/ora_probe/events on one conversation endpoint that poppy.json lists for the poppy protocol (an entry without access signed_in first, then the first in alphabetical order), when the file lists one. It carries no Session header, no token and no body. A correct site answers 400 with invalid_session in a JSON body (a 400 version_unsupported reads not tested, since Poppy-Version may be checked first, and so does a 403 extension_required; a 401 asking for an account at an entry with access signed_in reads not tested, since that entry needs account access first and the spec does not order the two checks). (T1-03)

  • POST to one MCP server that poppy.json lists (the first in alphabetical order), when the file lists one. It carries an MCP initialize request with no Session header and no token, skipped when the scan already saw this server answer 401. A correct site answers 401 with resource_metadata on every Bearer or DPoP challenge. A server that completes initialize with no token, refuses it before any token check with 400 invalid_session, 400 version_unsupported or 403 extension_required in a JSON body, or rejects it with the 400 and MCP error of a server that speaks only MCP 2026-07-28 has no 401 challenge to grade, so the row reads not tested. (T1-05)

  • GET to web.browser_session_endpoint in poppy.json, when the file lists one. It carries a plain GET of the join page, with no cookie, as a browser loads it; Ora never submits the code it carries. A correct site answers 200 with a form holding a code, a cookie that ties the code to the browser (SameSite Strict or Lax, and a __Host- name, since spec 6 says other hosts on the domain must not be able to set the cookie and only that prefix ensures it), Cache-Control: no-store, frame-ancestors 'none' (or X-Frame-Options: DENY) and a Cross-Origin-Opener-Policy that severs the opener. (T1-15)

  • POST to web.browser_session_endpoint in poppy.json, when the file lists one. It carries the join form with a made up code and no join cookie. A correct site answers an error status and no cookie (a CDN or load balancer cookie, or one the response clears, does not count). (T1-07)

  • POST to web.browser_session_endpoint in poppy.json, followed by /authorize, when the file lists one. It carries the same made up code as JSON, with no Session header and no token. A correct site answers 400 invalid_session (a 400 invalid_request, a 403 extension_required or a 400 version_unsupported reads not tested). (T1-16)

  • POST to the endpoint of the operations extension's version 3 entry in poppy.json, when the file lists one, as the last probe of the round, so it is sent only when the round has room under its cap of 6 requests. It carries a JSON body with a made up operation reference, with no Session header, no token and no extension declared. A correct site answers 400 invalid_session (a 400 invalid_request, which a reference the Session cannot use also gets, a 400 version_unsupported or a 403 extension_required reads not tested). (T1-17)

Ora does not yet send requests with a token or from an account that is signed in. For sign in, Session and conversation endpoints, Ora can check only that they refuse requests without one, not that they work.

Limits today

  • A probe round per target domain at most once every 6 hours.

  • At most 6 requests per round. A probe past the cap is not sent, and its row reads not tested.

  • Fetches of other domains the issuer lists in protected_resources are capped at 4 per host every hour.

  • A budget of 10 seconds for all PAP work in one scan.

Turning active probes off

A verified owner of a domain can turn active probes off in Ora Studio, under Identity layer, then Setup, for that verified domain. The Tier 0 reads of public files still happen. Anyone else can write to hello@ora.ai and we will stop them for your domain.

If we are causing you trouble

Write to hello@ora.ai. If you would rather we did not scan your site at all, say so and we will exclude the domain.