If you run a store and want to know whether a visitor claiming to be “Claude-Agent” is real, Anthropic now documents a two-step check: match the Claude-Agent token in the User-Agent, then confirm the source IP against a file Anthropic publishes at https://claude.com/.well-known/sail. The header alone proves nothing.

What the Claude-Agent token is

Per Anthropic’s support article, agentic browser traffic carries a real browser’s User-Agent plus the Claude-Agent token, added “to indicate that the browser is being driven by a hosted Claude agent.” The example string includes compatible; Claude-Agent/1.0.

Anthropic’s advice is to match on the token, not the full string, because other parts of the string can change.

It is a separate thing from Anthropic’s robots. The article names ClaudeBot, Claude-User and Claude-SearchBot as those, and says allowing or blocking agentic browser traffic only affects whether users can have Claude complete tasks on your site. Your crawler and fetcher settings are managed separately.

What the sail file says

We fetched claude.com/.well-known/sail on November 10, 2026. It is a TOML file that declares:

  • schema = "draft-kelley-webbotauth-sail-00", last modified 2026-10-22T18:36:00Z.
  • expires_after = 3600, in seconds, so the list is meant to be re-read hourly.
  • One prefix set, agentic-browser-use, with exclusive = true: all traffic from these ranges is documented in the file. It lists nine IPv4 addresses and no IPv6.
  • One service, Claude-Agent, with exhaustive = true: all egress traffic from this service originates from that prefix set.

Read literally, any request with the token from an address outside the file is not Anthropic’s agent.

How to verify a request

Anthropic’s support article gives the logic: “Anyone can send the same User-Agent header,” so confirm the source IP is on the published list before trusting the request. The addresses may change, so it recommends updating rules from the file automatically rather than hardcoding them.

  1. Match Claude-Agent in the User-Agent.
  2. Check the connecting IP against the current sail file.
  3. Treat a match on both as Claude’s hosted agent; treat the token alone as unverified.

The file format follows draft-kelley-webbotauth-sail-00, which Anthropic’s article notes is not yet a standard. Rules built on it may need rework if the draft changes.

Where this sits next to Web Bot Auth

The schema name ties the file to the Web Bot Auth family of drafts, but it is a different mechanism from signed requests. Web Bot Auth proves identity with a cryptographic signature on each request. The sail approach is lighter: a User-Agent token plus a published IP allowlist. Neither replaces the other today, and we found nothing in the sources above saying how they will relate.

What this means for a store

  • If your WAF or bot manager blocks unknown browsers, Claude’s agentic browsing may be challenged like any scripted client. A verified allowlist rule is the narrow way to let it through.
  • Your robots.txt rules for ClaudeBot or Claude-User do not govern this traffic, per Anthropic’s own description.
  • If you count AI traffic in analytics, a User-Agent that looks like ordinary Chrome will read as human unless you match the token.

FAQ

What is the Claude-Agent user-agent token?

It is a string, such as compatible; Claude-Agent/1.0, that Anthropic appends to a normal browser User-Agent when a hosted Claude agent is driving a browser on a user’s behalf. Anthropic recommends matching on the Claude-Agent token rather than the whole string.

How do I check that Claude-Agent traffic is genuine?

Compare the request’s source IP with the addresses Anthropic publishes at claude.com/.well-known/sail under the Claude-Agent service. Anthropic says anyone can send the same User-Agent header, so the IP check is what makes the identification trustworthy.

Does blocking ClaudeBot also block Claude-Agent?

No. Anthropic describes agentic browser use as different from its robots (ClaudeBot, Claude-User, Claude-SearchBot), and says those settings are managed separately from rules about agentic browser traffic.

Sources

An agent that can be identified is an agent a store can serve on purpose, which is the first step toward letting it complete a purchase.