CaraOne, the Cara One browser agent

Cara One is a health assistant app made by Cara Technologies, Inc. CaraOne is the browser agent behind it. When a signed-in member asks Cara One to do something on a website, such as book a flu shot, find an open appointment, or check a price, a browser in our cloud opens that site and does that one task for that one member.

• Every visit starts from a request by one signed-in member. The agent goes to the site the member chose, or to the practice, pharmacy or plan they asked about, and opens only the pages that task needs.
• It is not a crawler. It does not browse on its own, build a search index, or collect sites in bulk.
• It never sends a message, books, cancels or pays until the member has approved that exact step in the app.
• We also run a small number of test tasks with our own test account, to check the agent still works. They follow every rule on this page, and they never send, book or buy anything.

User agent

Every request carries the browser's own user agent (with "HeadlessChrome" kept) followed by this label. The product token for robots.txt is CaraOne.

CaraOne/1.0 (+https://www.caramedical.com/bot)

Signed requests (Web Bot Auth)

Every page and file request the agent sends to a site is signed with HTTP Message Signatures (RFC 9421), as Web Bot Auth describes. WebSocket connections are not signed. Each signed request carries a Signature-Agent header with the value "https://platform.caramedical.com", and Signature-Input and Signature headers (Ed25519, tag "web-bot-auth", valid for 60 seconds). The public key directory is:

https://platform.caramedical.com/.well-known/http-message-signatures-directory

A request that says it is CaraOne but has no valid signature from that key is not from us. Please tell us about it (see Contact).

Network address

Requests come from 54.89.199.116 (Amazon Web Services, us-east-1). We will update this page before that address changes.

CaraOne follows robots.txt (RFC 9309) on every page it opens, also when a member asked for that site by name.

• It reads /robots.txt for each site before it opens a page there, and uses the group for CaraOne (or CaraOne/1.0). When there is no such group, it uses the * group.
• It checks every page it opens or clicks to, and again the page a step lands on (after a redirect or a form). If that page is not allowed, it leaves without reading it.
• Scripts, images, styles and data requests that an allowed page loads by itself are not checked one by one, as in any browser.
• It keeps your file for up to one hour, so a change reaches it within the hour.
• If robots.txt answers with a 4xx status (such as 404), it reads that as no rules. If it answers with a 5xx status, a 429, or cannot be reached, it opens nothing on that site and tries again after 5 minutes.
• When a page is not allowed, the agent stops that part of the task. It does not try to reach the page another way, such as another address, a cached copy or another site. Cara One gives the member a link to finish the task themselves on their own phone.

To keep CaraOne off your site

Add this to your robots.txt:

To keep it off only some pages, for example booking pages:

CaraOne reads the Content-Signal lines in your robots.txt (see contentsignals.org), in the same group it uses for robots.txt.

• ai-input=no: the agent uses an AI model to read each page it works on, for the member's task. So when your site says ai-input=no, CaraOne opens no pages on your site, as if robots.txt disallowed them all. The member gets a link to do the task themselves.
• search: CaraOne never builds a search index and never shows your pages as search results.
• ai-train: Cara never uses your pages to train or fine-tune an AI model, whatever this signal says.

What happens to what the agent reads

Page content is used to do that member's task and to show that member what was done (for example, the time it booked). It is not sold or shared with anyone else. We keep short notes about how a site works, such as which button opens the booking form, with no member information in them, so the next task there takes fewer steps.

• One task runs one browser, much like one person. Tasks are short: most take under a minute of work on a site.
• At most 12 page opens a minute on any one site, across all members together.
• One task opens a page on a site at most once a second, and opens the same page at most 5 times.
• If a page answers 429 Too Many Requests, or 503 with a Retry-After header, every task leaves your site alone for the time you asked (at least 30 seconds, at most one hour, and 5 minutes when you did not say). The task does not retry that page by itself.
• If your site blocks the agent 3 times in an hour (a block page, a 403, a 429 or a captcha), it stays away for an hour.
• It never solves a captcha or other person check, and never tries to get around a block. If a site asks for a person check, the member may do it themselves, or the agent stops.

The quickest way to keep CaraOne off your site, or off some pages, is robots.txt (above). It takes effect within an hour.

If the agent causes a problem on your site, if you want it to work differently there, or if you would rather we use an official API, email hello@caramedical.com with "CaraOne agent" in the subject. Cara Technologies, Inc., 1266 Treat Ave, San Francisco, CA 94110.