Stop Pasting API Keys into AI Tools: Safer Options

Paper-cut illustration of a hand holding back a large key from a row of app windows, while a paper chain links an office building to one window beside a small dial set low
Paper-cut illustration of a hand holding back a large key from a row of app windows, while a paper chain links an office building to one window beside a small dial set low
Link the account, don't hand over the key. Illustrative artwork, not an Anthropic image.

Short answer: If a third-party AI tool asks you to paste in an API key, look for another way to connect it first. On 17 November 2026 Anthropic launched linked billing in beta. Tools that support it let you link your Claude Platform organisation instead of handing over a key: usage is billed to you, you set a monthly spend limit for that tool, and you can unlink it at any time. If a tool only takes a key, give it a key of its own with a hard spend limit, never your main one. One catch for UK firms: with linked billing, the tool's requests run in the tool's own Anthropic organisation, so your data-retention and IP allowlist settings don't apply to them.

This is a documentation review of Anthropic's announcement and developer documentation, read on 18 November 2026. I have not linked an organisation or tried any of the tools named here.

Why a pasted key is a problem

Anthropic's own announcement puts it plainly: an API key "works like a password: anyone who has the key can make requests billed to the customer". Once you paste it into someone else's tool, you're relying on that tool to store it safely. If it leaks from there, the bill still lands on you.

That isn't a theoretical risk. Google's Threat Intelligence Group said on 16 November 2026 that criminals are still stealing AI credentials and reselling access to AI models, which it calls "LLMJacking". Our guide to setting a hard spend limit on your AI API keys covers how to cap the damage. This page is about the step before that: not handing the key over at all.

What Anthropic's linked billing changes

Linked billing is for third-party tools that use Claude, such as coding agents and workflow automation apps. Instead of pasting a key into the tool, you click its link button and approve it on Anthropic's own sign-in page. According to Anthropic:

  • No key changes hands. "No API key changes hands, so platforms have no customer keys to store, encrypt or rotate," and your keys "never leave" your organisation.
  • An administrator approves it. Anthropic's developer documentation says the person approving needs the Admin or Billing role in your Claude organisation.
  • You set a limit for each tool. The approval page asks for a monthly spend limit for that tool, in whole US dollars. You can also set a default cap that applies to every linked tool without its own cap. When a tool reaches its cap, Anthropic refuses its requests until 00:00 UTC on the first day of the next month (midnight in the UK in winter, 1am in summer), or until you raise or remove the cap.
  • You can unlink at any time. Links and caps are managed on the Linked platforms page in the Claude Console.
  • It's a beta. Anthropic's documentation says the service is prerelease, that endpoints and behaviour can change, and that it "might not be suitable for production use".

Anthropic lists the tools that support it so far: Asana, ClickHouse, Factory, Gumloop, Harness, Higgsfield AI, Kilo Code, Linear, Major, Mintlify, Omni, OpenArt, Niteshift, Railway, Ramp, Runlayer, Tessera, Tines and Warp.

Four checks before you link a tool

  1. Check how you pay Anthropic. Only organisations that pay Anthropic directly for API use (prepaid credits or invoiced billing) can be billed this way. Anthropic says organisations billed through the AWS, Google Cloud or Azure marketplaces can't, and nor can organisations that have only a Claude subscription plan.
  2. Check which credit will pay. Anthropic says the monthly API credits that come with Claude Max, Ultra and Team plans can't pay for these requests, and nor can credits from its partner, VC and startup programmes. Credit bought in the Claude Console can.
  3. Check where your data settings stop. Anthropic says requests "run in the platform's organization", so the tool's data retention and encryption settings apply, not yours. Your own IP allowlist and zero data retention arrangements don't follow the request. If you'll send personal data through the tool, treat the tool as a supplier that processes it: check its data processing terms and add it to your records, as you would for any other processor. Our UK GDPR guide to AI agents and supplier due diligence checklist cover the questions to ask.
  4. Set a number for the spend limit. Once a tool is live, the approval page also offers Continue without a limit. Don't take it. Pick a monthly figure above your normal use, and review it after the first month.

If the tool only takes a key

Plenty of tools won't support linked billing yet. If you have to share a key:

  • Give the tool its own key in its own workspace. Never reuse the key your own apps run on. If it leaks, you revoke one key and nothing else breaks.
  • Put a hard spend limit on that workspace. The steps for Anthropic, OpenAI and Google are in our spend limits guide.
  • Paste it only into the tool's settings field. Not into a chat, a support ticket, an email or a shared document.
  • Ask how the tool stores it. "Do you encrypt customer API keys, and who on your team can see them?" is a fair question for any supplier.
  • Revoke it when you stop using the tool. A forgotten key in a tool you've cancelled is still a live key.

What this page doesn't cover

This page covers Anthropic's feature only. Before assuming your other AI providers have nothing similar, check their own documentation. A Claude Pro or Max subscription isn't an API key, so none of this applies to the Claude app on its own. For current Claude plan prices in pounds, see our Claude pricing guide and the UK AI price tracker. For the wider set of security habits for agents, see are agent skills safe?

Frequently asked questions

Is it safe to give an AI tool my API key?

It's a risk you can reduce, not remove. Anthropic says an API key works like a password: anyone who has it can make requests billed to you. If you must share one, give the tool its own key in its own workspace with a hard spend limit, and revoke it when you stop using the tool.

What is Claude linked billing?

It's a beta feature Anthropic launched on 17 November 2026. A supported tool lets you link your Claude Platform organisation instead of pasting in an API key. Usage is billed to your organisation, you set a monthly spend limit for that tool, and you can unlink it at any time from the Linked platforms page in the Claude Console.

Can I use linked billing with a Claude Pro or Max subscription?

No. Anthropic's documentation says organisations that have only a Claude subscription plan can't be billed this way, and nor can organisations billed through the AWS, Google Cloud or Azure marketplaces. The monthly API credits that come with Claude Max, Ultra and Team plans can't pay for these requests either.

Do my Claude data retention settings apply when I use linked billing?

No. Anthropic says the requests run in the tool's own organisation, so the tool's data retention and encryption settings apply, not yours. Check the tool's data processing terms before you send it personal data.

Sources

  • Anthropic, "Linked billing on the Claude Platform", 17 November 2026, opened 18 November 2026: https://claude.com/resources/articles/linked-billing-on-claude-platform
  • Claude Platform documentation, "Linked billing" (beta), opened 18 November 2026: https://platform.claude.com/docs/en/manage-claude/oauth-apps/organization-billing-connections
  • Google Cloud blog, "GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold", 16 November 2026, opened 18 November 2026: https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold/

Last checked: 18 November 2026 (Europe/London). Evidence: documentation review of Anthropic's announcement and developer documentation. I have not linked an organisation or tested any tool. General information, not security or legal advice.

Scroll to Top