claude-code.md

Claude Code

Claude Code speaks MCP natively. Wire AgentPwd in via the standard .mcp.json mechanism and the 10 tools become available to the agent.

Setup

After npm install + npm run build, drop this into your project's .mcp.json:

{
  "mcpServers": {
    "agentpwd": {
      "command": "node",
      "args": ["/absolute/path/to/agentpwd/dist/mcp/server.js"]
    }
  }
}

For user-level config (available across projects), use claude mcp add instead.

Initialize a vault before the first tool call:

node /path/to/agentpwd/dist/cli/index.js init
node /path/to/agentpwd/dist/cli/index.js add --site github.com --username me

Browser path

What AgentPwd will do depends on your OS and Chrome state:

OS	Chrome state	Path used
macOS	Already open (default)	AppleScript — enable Chrome → View → Developer → Allow JavaScript from Apple Events
macOS	Launched with --remote-debugging-port=9222	CDP, auto-detected (preferred when available)
Linux / Windows	Launched with --remote-debugging-port=9222	CDP, auto-detected
Linux / Windows	Already open without debug port	Not supported — use a remote browser via cdp_url

For the CDP path, launch Chrome like this and leave it running:

# macOS
/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome --remote-debugging-port=9222

# Linux
google-chrome --remote-debugging-port=9222

# Windows
"C:\Program Files\Google\Chrome\Application\chrome.exe" --remote-debugging-port=9222

Sample interaction

You: log me into github.com using my agentpwd credential
Claude: [calls ap_list_credentials → finds id]
        [calls ap_fill_login(credential_id="<id>")]
        → {status: "success", navigated: true}
        Logged in.

Optional hardening

Add a PreToolUse hook on mcp__agentpwd__ap_run if you want a confirmation prompt or a deny-list for shell commands the agent might try to run with credential injection. Example hooks ship in hooks/ (TODO).

See the threat model for what AgentPwd defends against and what it doesn't.