Skip to content

Claude-Agent: verify Anthropic's hosted browser by token and IP

How to verify the Claude-Agent user agent: parse Anthropic's SAIL file at claude.com/.well-known/sail, match token plus source IP, and write WAF rules.

By Crawlito · 8 min read

In short

  • Claude-Agent traffic is a real browser hosted by Anthropic, sending that browser's normal user agent with a Claude-Agent token added, and Anthropic says to match on the token.
  • Anthropic publishes Claude-Agent source IPs in a TOML file at claude.com/.well-known/sail, in the format of the SAIL Internet-Draft, not in the bots.json crawler list.
  • When read on 2026-11-15, the SAIL file listed nine single IPv4 addresses for Claude-Agent, no IPv6, marked exhaustive, with a one-hour lifetime.
  • Because the service is marked exhaustive, the SAIL draft treats a request that claims Claude-Agent from an address outside the list as refuted.
  • Anthropic's article does not say how Claude-Agent handles robots.txt, CAPTCHAs or bot challenges, so verify by token and IP and decide at your firewall.
On this page

To verify a request that carries the Claude-Agent user agent token, check its source IP against the Claude-Agent entry in Anthropic's SAIL file at https://claude.com/.well-known/sail. If the token is present and the IP is listed, the request is consistent with Anthropic's hosted browser. If the token is present and the IP is not listed, Anthropic's own file says it did not come from Claude-Agent. Fetch the file automatically, because Anthropic says the addresses can change.

The identity facts are short. According to Anthropic's Claude-Agent article, some Claude products, including Claude Duet, drive a real browser on Anthropic's infrastructure to complete tasks for a user. That browser sends its normal user agent with a Claude-Agent token added. It is separate from ClaudeBot, Claude-SearchBot and Claude-User, which are covered in ClaudeBot, Claude-SearchBot and Claude-User explained. This post is about the verification step: reading the SAIL file, matching token plus IP in a script and a firewall, and what a bot challenge does to one of these sessions.

What is in claude.com/.well-known/sail

The file is not JSON. When read on 2026-11-15 it came back with content-type: application/sail+toml and cache-control: public, max-age=3600, and its body was a TOML document:

toml
schema        = "draft-kelley-webbotauth-sail-00"
modified      = 2026-10-22T18:36:00Z
expires_after = 3600  # seconds

[[prefix_sets]]
name          = "agentic-browser-use"
exclusive     = true
ipv4_prefixes = [
  "136.71.21.147",
  "34.31.25.14",
  # ...seven more single addresses
]
ipv6_prefixes = []

[[services]]
name        = "Claude-Agent"
exhaustive  = true
prefix_sets = ["agentic-browser-use"]

The addresses above are a snapshot for illustration. Do not copy them into a config file. At the time of writing the list held nine single IPv4 addresses and no IPv6 prefixes, and none of the nine appeared in Anthropic's crawler list at claude.com/crawling/bots.json. An allowlist built from bots.json alone therefore does not cover Claude-Agent, and the reverse is also true.

Two flags in the file carry most of the meaning:

KeyWhereWhat it asserts
exhaustive = trueOn the Claude-Agent serviceAll egress traffic from Claude-Agent comes from the listed prefixes
exclusive = trueOn the agentic-browser-use prefix setTraffic from these addresses is documented in this file
expires_after = 3600Top levelThe document is valid for one hour after you receive it
modifiedTop levelWhen the content last changed

How the SAIL Internet-Draft format works

SAIL stands for Source Address Identity List. It is defined in draft-kelley-webbotauth-sail-00, an individual Internet-Draft by Kevin Kelley of Anthropic, dated 22 October 2026 and discussed on the IETF Web Bot Auth list. It is not a standard: the Datatracker page says it is not endorsed by the IETF and has no formal standing. Anthropic's article says the same, calling the format not yet a standard.

The rules a consumer needs, from the draft:

  • Location and transport. The file lives at /.well-known/sail over HTTPS only, with no fallback to plain HTTP. Consumers should follow redirects, but never to another scheme.
  • Structure. A UTF-8 TOML document of at most 512 KiB. Required keys are schema and modified, plus exactly one of expires (an absolute time) or expires_after (seconds, up to 30 days). [[prefix_sets]] group addresses; [[services]] name the clients and point at prefix sets or list prefixes directly.
  • Prefixes. CIDR notation or a bare address, which counts as a /32 for IPv4 or /128 for IPv6.
  • Names. Service and set names compare case-insensitively, so Claude-Agent and claude-agent are the same service.
  • Validity. Do not act on a document after its effective expiry, and never more than 30 days after you received it. Do not refresh a document that has not expired more often than every 1800 seconds. With Anthropic's one-hour lifetime, that means refreshing between 30 and 60 minutes.
  • Invalid files. A document that fails validation is ignored as a whole. Unknown keys are ignored.

The draft is strict about what the file lets you conclude. A request that claims a service and comes from that service's prefixes is a confirmation, and you may allowlist on it, keyed on the pair of claimed identity and source address, never on the address alone. A request that claims an exhaustive service from anywhere else is refuted. A request from an exclusive prefix that claims some other service is also refuted. The draft gives no basis for conclusions about traffic that claims nothing, so a request from a listed IP without the Claude-Agent token is not, by this file, proven to be anything.

The draft also positions SAIL as a complement to reverse DNS and to HTTP Message Signatures, the Web Bot Auth approach. Anthropic's article describes only the IP method for Claude-Agent, and does not mention signatures.

A script that matches token plus source IP

This script reads a downloaded copy of the SAIL file, collects every prefix for Claude-Agent, and checks the IPs you pass it. It uses only the Python standard library (tomllib needs Python 3.11 or later).

bash
curl -sS -A "my-site-verifier/1.0" https://claude.com/.well-known/sail -o sail.toml
python3 check_sail.py 34.31.25.14 203.0.113.9
python
import ipaddress, sys, tomllib

SERVICE = "claude-agent"  # SAIL names compare case-insensitively

with open("sail.toml", "rb") as f:
    doc = tomllib.load(f)
if doc.get("schema") != "draft-kelley-webbotauth-sail-00":
    sys.exit(f"unknown schema: {doc.get('schema')!r}")

sets = {s["name"].lower(): s for s in doc.get("prefix_sets", [])}
nets = []
for svc in doc.get("services", []):
    if svc["name"].lower() != SERVICE:
        continue
    raw = svc.get("ipv4_prefixes", []) + svc.get("ipv6_prefixes", [])
    for ref in svc.get("prefix_sets", []):
        ps = sets[ref.lower()]
        raw += ps.get("ipv4_prefixes", []) + ps.get("ipv6_prefixes", [])
    nets += [ipaddress.ip_network(p, strict=False) for p in raw]

print(f"modified {doc['modified']}, {len(nets)} prefixes")
for ip in sys.argv[1:]:
    addr = ipaddress.ip_address(ip)
    print(ip, "listed" if any(addr in n for n in nets) else "NOT listed")

Run against the file on 2026-11-15, it printed 9 prefixes, then listed for the first address and NOT listed for the second. Two practical notes from that test. The file is served through Cloudflare, and a request with Python's default Python-urllib user agent got a 403, so the example downloads with curl and a descriptive user agent. And the script deliberately fails on an unknown schema value: when the draft moves to a new version, check the new rules before trusting the output.

To find candidates in your logs first, match the token case-insensitively and print the client IPs. With the default nginx combined format:

bash
grep -i 'claude-agent' /var/log/nginx/access.log | awk '{print $1}' | sort | uniq -c | sort -rn

If your server sits behind a CDN or load balancer, the first field is the proxy's address, so log the real client IP before you run this. Then feed those IPs to the script. Any address that claims the token and comes back NOT listed is an impersonator. How to verify real AI crawlers covers the same pattern for other operators' lists.

WAF rules for Claude-Agent

The rule to write is the same whether you want to allow or block the real agent: decide on token and IP together. A rule on the token alone can be copied by anyone, and Anthropic's article says exactly that.

nginx

Generate the address list from the SAIL file on a schedule, then combine it with a user agent check:

bash
python3 -c 'import tomllib;d=tomllib.load(open("sail.toml","rb"));print("\n".join(f"{p} 1;" for s in d["prefix_sets"] for p in s["ipv4_prefixes"]+s["ipv6_prefixes"]))' > /etc/nginx/claude-agent-ips.conf
nginx -t && nginx -s reload

This shortcut takes every prefix set in the file, which is fine while the file has one service. If Anthropic adds services, use the full script's lookup instead.

nginx
geo $claude_agent_ip {
    default 0;
    include /etc/nginx/claude-agent-ips.conf;
}

map $http_user_agent $claims_claude_agent {
    default 0;
    "~*claude-agent" 1;
}

map "$claims_claude_agent$claude_agent_ip" $fake_claude_agent {
    default 0;
    "10" 1;   # claims the token, IP not listed
}

server {
    if ($fake_claude_agent) { return 403; }
}

To block the real agent instead, return 403 on "11". Keep the refresh job inside the draft's window: no more often than every 30 minutes, and before the one-hour lifetime runs out. If the fetch fails, the draft makes an exception to its expiry rule while the origin is unavailable, but says consumers using stale data in that case should exercise caution.

Cloudflare and other WAFs

In a Cloudflare WAF custom rule, the same logic is an expression on the user agent and the source IP, with the addresses held in an IP list that your refresh job updates through the API:

(lower(http.user_agent) contains "claude-agent" and not ip.src in $claude_agent_ips)

Set the action to Block for impersonators. For any WAF, the shape is the same: a managed IP set refreshed from the SAIL file, and one rule that requires both conditions. The general method for keeping operator IP lists current is in AI crawler IP ranges for firewall allowlists, and the equivalent for OpenAI's cloud browser is in how to allow ChatGPT's cloud browser through a WAF.

What a bot challenge does to a hosted-browser session

Anthropic's Claude-Agent article does not say how the hosted browser handles CAPTCHAs, JavaScript challenges or managed challenges. Its crawler article says ClaudeBot and the other crawlers respect anti-circumvention technologies and will not try to bypass CAPTCHAs, but that article says it covers Anthropic's crawlers and fetchers and sends readers to the separate Claude-Agent article for hosted browsers, so do not assume either way.

What you can say from how challenges work: a challenge replaces your page with an interstitial. Whatever the agent does with it, the user's task is held up on that step, and if the challenge is not passed, the agent never sees your content. WAF challenge pages and AI crawlers explains what each kind of challenge returns.

So make the decision explicit instead of leaving it to bot scoring:

  1. If you want users to be able to send Claude to your site, add a skip or allow rule for requests that match both the token and a listed IP, placed before your bot and challenge rules.
  2. If you do not, block the verified pair with a plain 403. Anthropic says that allowing or blocking this traffic controls whether users can have Claude complete tasks on your site.
  3. Either way, block or challenge requests that claim the token from unlisted IPs. Under the draft's exhaustive rule, their claim is refuted.

robots.txt is not a substitute here. Anthropic's Claude-Agent article does not say whether the hosted browser reads robots.txt, so a firewall rule is the only control whose effect you can confirm.

Frequently asked questions

Is Claude-Agent the same as Claude-User?

No. Anthropic covers Claude-User in its crawler article, with IPs in bots.json, and treats Claude-Agent as a separate hosted browser with IPs in the SAIL file. Write separate rules for each.

Does the browser use tool in the Claude API send Claude-Agent?

Anthropic's browser use tool documentation says the browser runs in the developer's own application, with nothing running on Anthropic's side, so sites see that developer's network identity. The page does not mention the Claude-Agent token, so do not expect that traffic to match Anthropic's SAIL addresses.

Why can't I find Claude-Agent IPs in bots.json?

Anthropic publishes them in a different file, claude.com/.well-known/sail, under the Claude-Agent service. When read on 2026-11-15, none of its addresses were in bots.json.

How often should I refresh the SAIL file?

The draft says not more often than every 1800 seconds for a document that has not expired, and Anthropic's file sets a lifetime of 3600 seconds. A refresh every 30 to 60 minutes fits both.

Can I trust a request from a listed IP that has no Claude-Agent token?

The SAIL draft makes no statement about traffic that does not claim a service, and says allowlist decisions should rest on the claimed identity and the address together. Treat it like any other unidentified request.

Check what an agent will find

Verifying Claude-Agent only decides who reaches your page. The free check shows what a single page returns to AI crawlers, with no account needed, including robots.txt rules and blocks at the edge.

Filed under AI crawlers. Topics: anthropic, ai-agents, server-logs, waf