Claude Agent SDK

Claude Agent SDK keeps its own auth hook. Authdog stays the identity provider. The planned sample attaches an agent access token there and refuses the run when that token is missing or userinfo rejects it.

Sample planned at claude-agent/python in authdog/agents-samples. That folder is not published yet, so this page does not pin claude-agent-sdk. Python 3.10+.

Channel auth, machine credentials, tool checks, and instructions for this framework are the other pages under Claude Agent SDK in AI.

The hook

async def require_authdog(input_data, tool_use_id, context):
    token = token_for_this_run()
    if subject_if_accepted(token) is None:
        return {
            "hookSpecificOutput": {
                "hookEventName": input_data["hook_event_name"],
                "permissionDecision": "deny",
                "permissionDecisionReason": "unauthorized",
            }
        }
    return {}

options = ClaudeAgentOptions(
    hooks={"PreToolUse": [HookMatcher(hooks=[require_authdog])]}
)

There is no matcher, so the hook runs for every tool. permissionDecision: deny stops that call. An empty object lets the normal permission flow continue. The token comes from the host (token_for_this_run), not from input_data["tool_input"]. The model does not get to supply it. Tests call the hook. They do not need ANTHROPIC_API_KEY.

This is agent identity: client_credentials, then this hook. It is not Agent Sign-In. A browser sign-in does not replace the hook.

What you should see

The hook returns permissionDecision deny when the token is missing or userinfo rejects it. The tool does not run.

The hook returns {} when userinfo accepts the bearer. A whoami tool may then return that subject.

pytest in claude-agent/python is how the sample will prove both. It does not call a model.

Configure

The shared exchange is in Agent frameworks. There is no refresh token. Repeat it before expires_in.

Related