1Password lets Claude sign you in without ever seeing your passwords

The News

1Password has switched on a browser integration that lets Anthropic's Claude sign in to websites on a user's behalf, the company confirmed on 16 July 2026. Branded "1Password for Claude", it lets the assistant work through multi-step jobs - checking a dashboard, redeeming credits or flagging unusual account activity - without the person typing a single username or password.

The twist is what Claude never touches. 1Password calls the design a "zero-exposure architecture": when the assistant hits a login wall, the password manager shows which credential is being requested and why, waits for a biometric approval, then injects the login and any one-time code straight into the page. The vault item, the password and the second-factor code never enter the model's context, the company says, and access is cut off the moment the job finishes.

"The answer isn't handing agents your secrets," said Nancy Wang, 1Password's chief technology officer. The system, she added, lets a user grant an agent permission to use a credential without letting the agent see it.

The integration is live now on Mac across 1Password's business, family and individual plans, and is bundled in for existing subscribers, with a 14-day free trial for new users.

Why It Matters

Autonomous AI agents have had one embarrassing gap since they learned to click around real websites: they are hopeless at the front door. When Anthropic first showed Claude operating a computer in October 2024, the demos stopped short of anything behind a login, because handing a probabilistic model your banking password is a security incident waiting to happen. Rival frameworks have mostly coped by asking users to paste credentials into a chat window - the exact habit security teams train people out of.

1Password's pitch is that delegation, not disclosure, is the fix. Instead of making the model trustworthy enough to hold a secret, the secret never moves. It echoes the logic that made "Sign in with Google" and passkeys palatable, now stretched to cover an agent acting on your behalf. If credential brokering becomes the control point for agentic commerce, whoever sits between a user's vault and the agent owns a valuable choke point.

Indian Angle

For Indian users and regulators, the friction is authentication law. The Reserve Bank of India's additional-factor authentication regime assumes a human clears each sensitive step, and its tokenisation and two-factor rules are among the strictest anywhere. An agent that completes a one-time code on a user's behalf, even without seeing it, sits awkwardly against that model. Any Indian bank or payments firm eyeing agentic logins must square the design with RBI expectations and with the consent rules under the Digital Personal Data Protection Act, 2023, which treats an agent acting for a user as processing personal data on their behalf.

There is a competitive angle too. Chennai-based Zoho ships Zoho Vault and has pushed its own Zia AI agents into Indian enterprises, giving it both the credential store and the assistant a design like this needs. If delegated-credential access becomes table stakes, Zoho and other Indian software firms are well placed to offer a home-grown version to businesses that already run their stack in rupees. India's vast developer base now has a concrete "let the agent act, never let it read the secret" pattern to build around.

FAQ

When is 1Password for Claude available?

It went live on 16 July 2026 for Mac, across 1Password's business, family and individual plans. Existing subscribers get it at no extra cost, and new users can try a 14-day free trial. Support beyond Mac was not detailed at launch.

Does Claude actually see my passwords?

No, according to 1Password. Its zero-exposure design injects the credential and any one-time code into the page after a biometric approval, so neither ever enters Claude's context. 1Password also checks the secret was not left exposed on the page.

How is this different from normal autofill?

Ordinary autofill assumes a human approves each login. Here an AI agent drives the browsing, and 1Password scopes access to a single task, revoking it once the job finishes rather than leaving credentials open for a whole session.

What should Indian fintechs watch for?

Whether agent-completed logins and one-time codes fit the RBI's additional-factor authentication norms and the consent duties under the DPDP Act, 2023. Firms should treat an agent as a data processor acting for the user and build audit trails accordingly.

This story was reported by The Verge. Read the full original coverage at The Verge.