Agent frameworks

When your agent calls a tool, it needs a token your API already knows how to check. Authdog issues that token. LangGraph, the OpenAI Agents SDK, the Claude Agent SDK, and A2A all send the same bearer, and you verify it on your server.

The Agents module is sales-gated. Until that access is approved, create a machine-to-machine application and use client_credentials. Keep a development token off your production API.

What the token says

Reject a bad or missing token with 401. Reject a valid token that lacks the scope or tool with 403. Check issuer, audience, and expiry on every call.

LangGraph

Give the worker the client id and secret. Exchange them before the graph runs, and attach the access token on each tool request.

curl -X POST "$TOKEN_ENDPOINT" \
  -u "$AUTHDOG_CLIENT_ID:$AUTHDOG_CLIENT_SECRET" \
  -H "Content-Type: application/x-www-form-urlencoded" \
  --data-urlencode "grant_type=client_credentials"

Pass Authorization: Bearer <access_token> from the tool node. Refresh when expires_in is close. Do not put the client secret in the prompt.

OpenAI Agents SDK

Register one confidential client per agent. In the tool's execute, read the bearer your channel already verified and call your API with it. A tool that needs a person's account uses token exchange so the user stays sub and the agent is in act.

Claude Agent SDK

Same exchange. Put the access token on the MCP client or HTTP tool, not in the system prompt. If a tool is marked as requiring approval, a 403 with reason approval_pending means wait. Do not retry until an operator approves it in Agents → Approvals.

A2A

Put this on the Agent Card. The client fetches the token out of band and sends it on the HTTP request, not inside the JSON-RPC body.

{
  "securitySchemes": {
    "authdog": {
      "openIdConnectSecurityScheme": {
        "openIdConnectUrl": "https://<issuer>/.well-known/openid-configuration"
      }
    }
  },
  "securityRequirements": [{ "authdog": ["mcp:invoke"] }]
}

Related