1Password for Claude: Delegating Sign-In to an AI Without Showing It Your Passwords
On 16 July 2026 (17 July in Japan), the password management service 1Password announced 1Password for Claude, an integration with Anthropic. It lets Claude sign in to websites in a browser and carry on working.
More important than the convenience is that Claude is never shown the password. The user approves each time, and 1Password enters the credentials and one-time codes into the page directly. Claude learns that the sign-in succeeded, but the secret strings themselves enter neither the model’s context nor its memory.
The problem with delegating sign-in to an AI
An AI agent operating a browser can compare products, make reservations, check account information, and pull figures from an admin panel. Reaching a sign-in page, however, stops the work.
Having a human sign in instead is safe, but means taking the work back each time. Typing a password into the chat, or storing it somewhere the AI can read, is dangerous. 1Password for Claude fills that gap by granting permission to use a credential for this job rather than disclosing the secret.
How a sign-in proceeds
- Claude is working in the browser and reaches a page requiring sign-in
- Claude requests the necessary credential from 1Password, with the purpose
- An approval screen appears in 1Password’s Mac app; the user selects the item and authorizes with biometrics or similar
- 1Password enters it into the browser directly, and Claude resumes once submission completes
Beyond ordinary sign-in fields, TOTP one-time passwords are supported. If a sign-in fails, the entered values are cleared before Claude reads the page again.
“Zero Exposure” keeps secrets out of the model
1Password calls this design Zero Exposure. What reaches Claude is the name of the item the user approved, the username or email address, the registered website, and whether the operation succeeded or failed. The password and the one-time code themselves do not.
Access is scoped per Claude session rather than granted as a standing blanket permission; a new session requires approval again. Communication containing secrets happens inside the Mac, and approved information is discarded when the session ends or after a time limit.
Agentic Mode hides the vault itself
The other mechanism is Agentic Mode in the 1Password browser extension. While an AI agent is operating the browser, it hides 1Password’s suggestion and save interfaces from the page.
This stops an agent from clicking through 1Password’s own interface and reaching items it was not granted. All Claude can use is the credential the user explicitly approved for that job. Agentic Mode is designed to work with other supported agents too.
Available on Mac; four things required
1Password for Claude is available on Mac for Individual, Families, and Business plans. It requires:
- The 1Password desktop app for Mac
- The 1Password browser extension
- The Claude desktop app
- The Claude in Chrome browser extension
With the apps and extensions installed, you pair Claude and 1Password once at the start. 1Password verifies the app’s signature and confirms it is a genuine Anthropic app before connecting.
What is made safe is the handover of credentials
Worth noting is the scope of what 1Password guarantees. Storing the password, approving it, and entering it into the page are protected. What Claude then does on the website is left to Claude’s own safety features and to your instructions.
Signing in safely to an admin panel does not guarantee that the deletion, purchase, publication, or settings change that follows is what you intended. For operations that matter, permission to sign in is not enough; confirmation before execution and limits on scope are needed too.
Cutting a spare key for the age of AI agents
Password managers have been safes for humans to sign in from. With 1Password for Claude, the secret never leaves the safe; what the AI receives is a short-lived spare key usable only for the job at hand.
The more AI agents move from thinking to booking, buying, and administering, the more the question shifts from what they know to which permissions they hold, on whose approval, and for how long. 1Password for Claude is a clear demonstration of the direction: manage permissions rather than teaching an AI your passwords.
References: 1Password, “1Password for Claude: Give Claude access without giving up your credentials”; 1Password Support, “About the security of 1Password for Claude”. Supported environments and specifications are as of July 18, 2026.
