AI agent credential tools fall into two camps in October 2026: those that let an agent sign in to a website without ever holding the password, and those that keep API keys and tokens out of the agent’s code. 1Password for Agents, announced on October 22, 2026, is the clearest consumer example of the first camp, and it is the news that prompted this comparison.

According to 1Password, when an agent reaches a login page it can request a credential from your vault. You review and approve the request, and 1Password fills the login on the site, so the agent never gets the password or the one-time code. The product costs nothing extra on Individual and Family plans on 1Password.com, though accounts on 1Password.eu and 1Password.ca are not currently supported. This report compares it with five other options and says which job each one does.

Two different problems behind one headline

A person asking an assistant to book a table or renew a subscription needs it to log in to a website with a username and password. A developer building an agent that reads a calendar or posts to Slack needs it to call an API with a token. The first is a human-credential problem with an approval step. The second is a machine-credential problem of storage, rotation and scoping. 1Password for Agents and Bitwarden’s Agent Access SDK sit in the first group. Auth0 Token Vault, Infisical Agent Proxy and Keeper Secrets Manager sit mostly in the second. Dashlane has described the problem but, as of its February 2026 post, not shipped a dated agent-delegation feature.

The design choice that matters: does the agent ever see the secret?

The sharpest difference across these tools is whether the plaintext secret ever enters the agent’s context. 1Password’s launch design fills the credential into the page itself. Bitwarden’s SDK describes injecting the credential into the agent process without plaintext exposure, after a human approves each request. Infisical’s proxy gives the agent only a placeholder and attaches the real key outside the agent. Keeper’s MCP integration, by contrast, can return secret values to the assistant after confirmation, and its documentation notes that the third-party AI tool’s own policies govern that data. Which design is acceptable depends on how far you trust the agent and the model provider behind it.

What to check before relying on the launch

The 1Password announcement is days old. Its own post says availability, platforms, capabilities, safeguards and controls vary by partner integration. Its list of agents mixes products rolling out now with ones marked coming soon, and a second report describes some of them as already connected. To find out whether your agent works today, check that agent’s own settings. 1Password and Anthropic launched an integration for Claude on July 16, 2026, reported at the time as a beta, so this launch extends an existing approach.

Where this is heading

Dashlane’s CEO pointed to the lack of fine-grained permissions for agents, and Bitwarden’s Agent Access SDK is published as an open protocol rather than a closed feature. Together with 1Password’s partner list, these point toward a per-request approval step in which the vault, not the agent, handles the secret. Nothing published so far settles whether these converge on one standard or stay vendor-specific.