axsDocs
axs cred

Agents and credentials

How an agent logs in and runs commands without the values.

Log in to a site

The MCP server axs-cred gives the agent two tools:

  • cred_list: the credentials it may use, with name, sites, login and whether a TOTP is stored. No secret values.
  • cred_fill: fills login, password and, when the page asks, the TOTP code into the open tab of the credential's site and submits. It returns only what was filled.

After a fill, the agent cannot read that tab's page until it navigates.

From another machine

An agent on a machine that is not enrolled (a dev box, a VM, with Claude in Chrome driving your Chrome) needs only axs login and axs cred setup there. Its cred_list and cred_fill go to one of your enrolled devices whose Chrome runs the axs cred extension: that device decrypts and fills its own Chrome, and the values never leave it. Approvals work the same: type approve <code> in the agent's chat.

Agent browsers

An agent can also log in inside a browser it drives itself: Playwright MCP, a Playwright script or any Chrome with a DevTools port. Turn on Agent browser under Options for each credential that may be filled there; the others are refused.

cred_fill looks for a tab on the credential's site in the agent browser first and falls back to Chrome. It reaches the browser on 127.0.0.1:9333, then 9222; set AXS_CRED_CDP=host:port or pass cdp_url for another port. Only local ports are used.

  • Playwright MCP: axs cred setup adds the DevTools port to your Playwright MCP. axs cred setup --playwright adds a Playwright MCP if you have none.

  • Playwright scripts: launch the browser with --remote-debugging-port=9333, open the login page, then

    axs cred fill github

    --no-submit fills without submitting, --cdp host:port picks the browser.

Use a value in a command

In a shell command the agent writes a placeholder:

curl -u admin:{{cred:grafana/password}} https://grafana.example.com/api/health

The hook turns it into axs cred run, so the value reaches only that command, never the transcript or the command line. Fields are login, password and totp.

You can run it yourself too:

axs cred run --map PGPASSWORD=prod-db/password -- psql -h db.example.com -U app

Values in the command's output are masked.

Leaks

The hooks deny tool calls whose input carries a stored value, and strip stored values (also base64, URL encoded, hex or reversed) from tool output. A credential that showed up in tool output is locked: change its password in the dashboard to unlock it.

Approvals

Each credential has a mode:

ModeEffect
AutoThe agent uses it without asking
Ask onceYou approve once, then it works for a while (15 minutes to 7 days)
Ask every timeYou approve each use

When a credential asks, the agent tells you a code. Type approve <code> in the chat yourself; the code is valid for 10 minutes. Without the server a credential that asks is refused.

On this page