# Claude Console IP allowlisting: what to pin to your gateway

URL: https://infro.io/blog/claude-console-ip-allowlisting-what-to-pin-to-your-gateway
Author: INFRO Editorial (https://infro.io/blog/editorial)
Publisher: INFRO
Published: 2026-11-07
Last modified: 2026-11-07
Last reviewed: 2026-11-07 (next review due 2026-12-07)
Topic: Security and governance
Format: Checklist

> List only the public egress addresses of your gateway and the few human or CI networks that truly need Claude access, then prove a direct call from anywhere else returns a 403 permission_error. The list is org-wide, takes up to 15 minutes to apply, and does not cover Bedrock, Vertex or Foundry. [^release-notes][^ip-allowlist]

Key takeaways:
- The Anthropic allowlist is one list for the whole organization, covering the Console, the main API endpoints and the Admin and Compliance APIs, so every legitimate network must be on it.
- A gateway's egress addresses are the entries that matter most; if the gateway fails over across regions or hosts, each egress address needs an entry.
- A gateway that retries on 403 can hide a missed allowlist entry behind a slightly slower response, so watch for upstream 403s rather than client errors.
- The allowlist does not apply to Claude on Bedrock, Claude Platform on AWS, Google Cloud or Foundry, so failing over to those paths leaves it behind.
- Evidence for a reviewer is the entry list with an owner per entry, a recorded denied direct call, and the change history.

## What changed and what it covers

On November 5, 2026 Anthropic announced that you can set up an IP allowlist for your organization in the Claude Console, in beta. Requests to covered services from addresses not on the list are refused. Entries are IPv4 and IPv6 addresses or CIDR ranges, up to 1,000 per organization.[1]

The list is organization-wide. Anthropic compares each covered request's public IP address with it and denies the request if the address is not allowed. The Messages API returns a 403 `permission_error` with the message "Your IP address isn't in this organization's allowed range."[2]

For a Console organization, the list covers:

- the Console UI;
- the Messages, Message Batches, Token Counting, Models, Files and Skills endpoints, with an API key or a Workload Identity Federation token;
- the Claude Managed Agents endpoints for Agents, Sessions, Environments, Vaults and Memory stores;
- the Admin API and Compliance API, with a key that covers only this organization.[2]

The goal of this checklist is narrow: Anthropic keys that work only from networks you can name, with a gateway as the main one. Per-end-customer restrictions are a different control, covered in [Anthropic profile restrictions: enforce them at the gateway](https://infro.io/blog/anthropic-profile-restrictions-enforce-them-at-the-gateway).

> **Warning: Beta**
> Anthropic states that settings and behavior can change during the beta.[2] Re-read the page before each review cycle and date your evidence.

## Decide what belongs on the list

Because the list is one set for the whole organization, "what to pin to your gateway" really means "what is the smallest set of networks that must reach Anthropic". Build it from an inventory, not from whatever address you are on today.

- [ ] List every workload that holds an `sk-ant-` key or a federated token, and the public address it leaves from.
- [ ] Mark which of those should go through the gateway instead, and plan to move them before allowlisting.
- [ ] Add the gateway's public egress addresses as entries, one per address or tight CIDR.
- [ ] Add only the CI, office or VPN networks that truly need direct Claude access, including Console users.
- [ ] Give each entry an owner and a purpose in your inventory, since Anthropic's list has no such field that I read.
- [ ] Use public addresses only; private ranges such as 10.0.0.0/8 are not what Anthropic sees.[2]
- [ ] Respect the format limits: no broader than /8 for IPv4 or /32 for IPv6.[2]

Anthropic also warns about shared proxies: if you list a shared proxy address, an attacker with a leaked key could potentially use it through that proxy.[2] That applies to any hosted gateway whose egress addresses are shared across customers. Ask the gateway vendor whether egress addresses are dedicated to you. The INFRO docs pages read for this article do not list egress addresses, so confirm them with INFRO before you pin anything; do not copy addresses from a blog post.

> **Example: Illustrative inventory**
> These entries use documentation address ranges and invented owners; they are not real addresses.
> 
> | Entry | Purpose | Owner | Review date |
> | --- | --- | --- | --- |
> | 203.0.113.10/32 | Gateway egress, region A | Platform | 2026-12-01 |
> | 203.0.113.74/32 | Gateway egress, region B | Platform | 2026-12-01 |
> | 198.51.100.0/28 | CI runners for eval jobs | ML platform | 2026-12-01 |
> | 192.0.2.0/29 | Office VPN, Console admins | Security | 2026-12-01 |

## Gateway egress and regional failover

A gateway changes which address Anthropic sees. Your applications talk to the gateway; the gateway calls Anthropic with your key. So the entries that matter are the addresses the gateway calls from, not your application servers.

Regional failover is where lists go stale. If the gateway has more than one region, or fails over between hosts, each path can leave from a different public address. A list that holds only the primary region works until the primary region has a bad day, which is when you need the second one.

- [ ] Get the full set of egress addresses, including those used only for failover, from the gateway vendor in writing.
- [ ] Ask how and when that set changes, and who tells you before it does.
- [ ] Prefer addresses you can verify yourself (a dedicated address or a NAT you own) over ones you must take on trust.
- [ ] Re-run your failover drill after the allowlist goes live, not before. See [Failover drills](https://infro.io/blog/llm-failover-drills) if you have no drill yet.
- [ ] Remember Anthropic says changes can take up to 15 minutes to apply, so add new addresses ahead of a planned move.[2]

There is a second trap. INFRO's failover moves to the next connection on every upstream failure before the first token, including a rejected credential (401/403).[3] A gateway that behaves that way will turn an Anthropic 403 from a missed allowlist entry into a retry on another connection, and the client sees only a slightly higher time to first token.[3] That is good for availability and bad for noticing a wrong list. Watch for upstream 403s per connection, not only client errors.

One more consequence: the Models endpoint is covered by the allowlist.[2] Any job that lists models with your key, such as an availability check, needs a listed address too.

## What the allowlist does not cover

Read the Limitations section before you tell a reviewer the key is "locked down". Anthropic states that the allowlist does not apply to Claude in Amazon Bedrock, Claude Platform on AWS, Claude Models on Google Cloud, or Claude in Microsoft Foundry.[2] It also covers requests to the Managed Agents endpoints, not the agents themselves, and does not apply to any networking in the Managed Agents runtime.[2]

That matters for failover design. If a route fails over from the Anthropic API to Claude on a cloud platform, the second path is governed by that cloud's identity and network controls, not by this list. Record that in your design, and put matching controls on that side.

- [ ] Name every non-Anthropic-API path to Claude that your routes can reach.
- [ ] For each one, record which control restricts it (cloud IAM, private networking, or none).
- [ ] State in the review that the Anthropic list governs only the Anthropic API and Console.

Also keep in mind the allowlist bounds where a key works, not how long. Anthropic recommends a short key expiration or Workload Identity Federation for that.[2] Key expiry is chosen when the key is created.[4]

## Block direct calls, without locking yourself out

Direct-call blocking is the point of the exercise: a key copied from a laptop, a repo or a build log should fail from any address you did not list. The mechanism is the allowlist itself, and it works only if the list is tight. Every extra office, home VPN exit or CI range is a place a leaked key still works. For egress-side blocking inside your own network, see [How to block direct AI provider calls with egress rules](https://infro.io/blog/how-to-block-direct-ai-provider-calls-with-egress-rules).

Lockout is the other failure. Anthropic notes that saving the list checks your current address but not your servers, teammates or tomorrow's address, and that organization admins are blocked too.[2]

- [ ] Enable the allowlist toggle only after every legitimate network is entered; nothing changes until you save.[2]
- [ ] Decide deliberately on "Allow admins to access this settings page from any IP address". Anthropic recommends it to avoid lockout, but a compromised admin session could then edit the list from anywhere.[2]
- [ ] If you take that option, compensate: two-factor on admin accounts and alerting on list changes.
- [ ] Watch for 403 `permission_error` responses after the change, including jobs that run only daily or weekly.[2]
- [ ] Do not allowlist home or mobile addresses; Anthropic notes they change without notice.[2]

A direct-call test belongs in the runbook. Run it from a host that is not on the list, with a throwaway key scoped to a test workspace:

```python
import os, anthropic

client = anthropic.Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"])
try:
    client.messages.create(
        model=os.environ["TEST_MODEL"],
        max_tokens=16,
        messages=[{"role": "user", "content": "ping"}],
    )
    print("FAIL: request was allowed from an unlisted address")
except anthropic.PermissionDeniedError as e:
    print("PASS:", e.status_code, e.message)
```

Expect a 403 whose message says the address is not in the organization's allowed range. Save the output with the date and the source address.

## Add the gateway's own key controls

The Anthropic list is the provider-side layer. Your gateway keys are the other. Applications hold gateway keys, never the provider key, and a gateway key can carry its own network limit. INFRO's authentication docs describe a gateway key called from outside its allowed networks returning 403 with code `ip_not_allowed`.[5] INFRO lists limiting a gateway key to IP ranges as a key control, so a copied gateway key is bounded as well as the provider key behind it.

Two layers fail differently, which helps you read logs:

| Symptom | Where it was refused | First place to look |
| --- | --- | --- |
| 403 `ip_not_allowed` from the gateway | Gateway key network limit | Calling host's address vs the key's ranges[5] |
| 403 `permission_error` from Anthropic | Provider organization allowlist | Gateway egress address vs the list[2] |

For the second row, the gateway's per-request record is what tells you which connection saw the 403 before failover moved on. INFRO's request traces and per-request usage record the connection that served each request, which helps find a connection whose egress address is missing from the list.

## Evidence to give the reviewer

Reviewers need proof that the control exists, is tight, and was tested. See also [AI access review checklist: evidence to export for model keys](https://infro.io/blog/ai-access-review-checklist-evidence-to-export-for-model-keys).

- [ ] The current entry list, dated, with owner and purpose for each entry.
- [ ] The gateway vendor's written egress address set, including failover addresses, and your check that every one is on the list.
- [ ] The recorded denial from an unlisted host (the script output above) and a recorded success through the gateway on the same day.
- [ ] The admin-bypass decision: whether "any IP" is on for the settings page, and the compensating controls.
- [ ] Change history. Anthropic says the Compliance API Activity Feed records changes to the allowlist, if you have set the Compliance API up.[2] Gateway-side, keep the audit log of key and policy changes.
- [ ] The key inventory with expiry dates; the Admin API can list API keys for auditing.[6]
- [ ] The scope statement: Anthropic API and Console covered; Bedrock, Claude Platform on AWS, Google Cloud and Foundry not covered.[2]
- [ ] The beta note and the date you last re-read Anthropic's page.

Finally, a list is a point-in-time control. Plan a quarterly check that every entry still has an owner, and a post-change check each time the gateway vendor announces an address change. If the gateway is down, what traffic does next is a separate decision; see [Gateway down: should AI traffic fail open or fail closed?](https://infro.io/blog/gateway-down-should-ai-traffic-fail-open-or-fail-closed). Direct calls that skip the gateway entirely are covered by [Detect AI API usage that bypassed your gateway](https://infro.io/blog/detect-gateway-bypass-provider-usage).

## Questions

### Does the Anthropic allowlist apply per workspace or per key?

Per organization. Anthropic describes one list of IP ranges that covers your entire organization, so you cannot give one workspace or one key a different list. [2]

### Is the allowlist enough if a key leaks?

No. Anthropic says an allowlist limits where a key works but not for how long, and recommends a short key expiration or Workload Identity Federation as well. [2]

## Sources

1. Claude Platform release notes — Anthropic. https://platform.claude.com/docs/en/release-notes/overview (accessed 2026-11-07)
2. IP allowlisting — Anthropic. https://platform.claude.com/docs/en/manage-claude/ip-allowlist (accessed 2026-11-07)
3. Failover & fallbacks — INFRO. https://infro.io/docs/features/fallbacks (accessed 2026-11-07)
4. Authentication — Anthropic. https://platform.claude.com/docs/en/manage-claude/authentication (accessed 2026-11-07)
5. Authentication — INFRO. https://infro.io/docs/authentication (accessed 2026-11-07)
6. Admin API — Anthropic. https://platform.claude.com/docs/en/manage-claude/admin-api (accessed 2026-11-07)
