1Password for Claude lets Claude complete browser logins without putting your passwords or one-time codes inside the AI model.
This is the right way to give an AI agent access.
Browser agents can already compare products, fill forms, and complete purchases.
But they often stop at the most important part: logging in.
You can give the agent your password, which is a bad idea. Or you can take over every time, which breaks the workflow.
1Password for Claude adds a third option.
How 1Password for Claude works
Claude never receives the password.
When Claude reaches a login page, 1Password shows you which credential it wants and why. You approve the request with Touch ID.
1Password then fills the credential directly into the page. The password and one-time code never enter Claude's context or memory.
The access only applies to the current task. It ends when the task ends.
1Password also checks that the page did not expose the secret after filling it. If the form submission fails, it clears the values before returning control.
This is a small detail with a big effect.
Claude knows it logged in. It does not need to know how it logged in.
Agentic Mode protects the rest of your vault
There is another problem to solve.
A browser agent can control the same browser where the 1Password extension lives. Without extra protection, the agent might try to interact with the extension itself.
1Password calls its solution Agentic Mode.
When a compatible agent takes over the browser, the extension locks down. Its interface becomes hidden, and the agent can only use credentials you approved for that task.
The rest of the vault stays unavailable.
Agentic Mode also works when you are not using the Claude integration. It protects 1Password when other compatible browser agents take control.
Why this matters for AI agents
We often treat agent access as a permissions problem.
It is also a secrets problem.
An agent may need permission to use Stripe, GitHub, or an admin dashboard. That does not mean the underlying credential should become part of the prompt.
My advice is to separate using a credential from seeing a credential.
We already do this in other parts of software. A deployment workflow can receive short-lived access without giving every process a permanent cloud key.
Browser agents need the same model:
- access is explicit
- access is scoped to one task
- the user approves it
- the secret stays outside the model
- access disappears when the work ends
This gives the agent enough power to finish the task without handing it every key.
What you need
1Password for Claude is available now on Mac for individual, family, and business plans.
You need:
- the 1Password desktop app
- the 1Password browser extension
- the Claude desktop app
- the Claude in Chrome extension
Once connected, Claude can use logins already stored in 1Password. You still approve each credential request.
The bigger idea
AI agents are becoming users of our software.
They click buttons, call APIs, edit files, and make purchases. They need identities and permissions, but they should not inherit unlimited access from the person running them.
1Password for Claude is one product integration. The more important part is the pattern.
Give the agent the capability it needs.
Keep the secret out of reach.