1Password launches platform where AI agents sign in without ever seeing the passwords
1Password launched 1Password for Agents on October 22, 2026, a platform that lets personal AI agents complete website sign-ins without seeing or storing the passwords. The user approves each request on their own device, and 1Password fills the credential directly into the form — an approach the company itself describes as a "trust layer" for agents, but which so far is documented only through the company's own descriptions.
What was launched
In a press release via Business Wire, 1Password announced that the new platform is meant to help personal AI agents "use the passwords, secrets and credentials they need to complete approved tasks without seeing or storing them" (b5b354ca). At its core is a principle that differs from both traditional delegation and most existing integrations: the agent is never given access to the password value itself.
How the mechanism works
According to the press release, the sign-in happens in three steps (b5b354ca):
- When the agent reaches a login form, it verifies its identity to 1Password with a unique digital key — one per agent.
- The user sees which credential the agent needs and why, and chooses whether to approve the request. The approval happens on the user's own device, which is intended to confirm that the permission actually comes from that person.
- 1Password then fills the credential directly into the login form, instead of sending it to the agent.
The Verge (Jess Weatherbed) describes the flow as reminiscent of two-factor authentication: the app sends a notification about which account logins are needed and why, and if the user approves, 1Password enters the information itself (59db214a).
"Agents have started doing real work on our behalf. The old model where you hand over a credential and hope for the best doesn't work," said Aashish Tripathi, chief product officer (CPO) at 1Password, in the press release (b5b354ca) (quote translated). The company describes the platform as a "trust layer" for agent experiences.
Rollout and who can use it
The first partner integrations are rolling out in the days after launch: Meta's Muse assistant and Grok Bot are set to be the first connections. Integrations for Anthropic's Claude Duet, OpenAI's ChatGPT Work/dots and Towns' Townie assistant are promised "soon," with Townie support planned for web and iOS (13be4089).
In addition, 1Password lists a longer roster of partners working on support: Browserbase, Browser Use, Dana, Instinct, Kernel, Perplexity, Ramp, Tomo and Wajo, with "more integrations on the way" (59db214a).
Availability has clear limits. The feature is included at no extra cost for Individual and Family accounts hosted on 1Password.com. Accounts on 1Password.eu and 1Password.ca are not supported (13be4089; 10db3051). 1Password is priced from $3.99 per month with annual billing (10db3051).
Background: the July Claude integration
This is not 1Password's first step into agent sign-ins. On July 16, the company and Anthropic launched a Claude integration that runs through Claude's desktop app and the Claude in Chrome extension on Mac. The setup requires a paid Claude plan and version 8.12.28 or later of both the 1Password app and the browser extension (13be4089). The new platform is thus a generalization of a concept first tested with a single partner, now opened to a broader ecosystem.
The contrast with Proton Pass
Competitor Proton has chosen a different model for giving AI tools access to passwords. Proton Pass offers expiring access tokens: the user chooses which vaults the token can read, sets an expiry from one hour to one year, and can delete the token to stop further access. AI access requires a stated reason, and all accesses are logged (13be4089).
The difference is significant: in Proton's model, the agent can actually read the password values — the control lies in time limits, scope and logging. In 1Password's model, the agent is never supposed to see the value at all, but in exchange every sign-in requires active user approval. Proton's model is better suited to autonomous agent missions without a user in the room; 1Password's model places responsibility for each individual sign-in with the human. Neither model has been independently tested in the available source material.
For the partner program, 1Password's chief business officer (CBO) John Torrey told The Deep View that integrations must uphold shared security principles and pass a security review (13be4089).
Caveats and open questions
Several caveats should travel with the reader. First, the entire mechanism description rests on 1Password's own account in the press release; none of the available sources have tested the platform in practice, and there is as yet no independent technical documentation of how the digital key and the form-filling work in detail.
Second, the protection only applies up to sign-in. WinBuzzer points out that what the agent can do after logging in depends on each agent's own security mechanisms — 1Password for Agents solves the credential problem, but not the question of what an agent does inside a logged-in session (13be4089).
Third, the details of the partner security review have not been published, and the criteria behind Torrey's statement are unknown. Several of the partner names in the launch material — including "SpaceXSI's Grok Bot" and a Muse quote attributed to Meta Superintelligence Labs — are unusual and unverified in the public source material, and should be treated with caution until 1Password confirms them.
What remains after launch week is a clear product philosophy: 1Password is betting that the future of agent sign-ins should look as much as possible like today's two-factor flow — a notification on your device, a deliberate choice, and a password that never leaves the password manager. Whether the model holds up when agents need to carry out long, autonomous missions is among the most important open questions in the weeks ahead.