Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Zalando's broker hands agents OAuth safely

Written November 8, 2026

Zalando open-sourced an inspectable OAuth delegation and consent broker for AI agents. What it brokers, and what it doesn't solve.

The problem it answers

Handing an agent an OAuth token means handing it the keys to a third-party service, with no inspectable record of what the agent was allowed to do with them. The token travels with the agent, the agent calls whatever it calls, and the only record is whatever the agent chose to log. Zalando's incubator open-sourced the Agentic Identity Broker to move that arrangement somewhere auditable: the agent never holds the user's provider token at all.

The broker is MIT-licensed and written in Go. Its repository describes it as managing delegation chains for on-behalf-of flows, bridging distinct OAuth 2.0 infrastructures, and providing a token vault. Zalando says it runs the broker in production internally for several MCP servers.

What it brokers

The delegation and consent model centres on three objects:

The call path runs through a gateway. The agent sends its own user-agent token only to the trusted gateway — Zalando's post names agentgateway — and the gateway validates that token and asks the broker to exchange it for an access token issued by the target's OAuth 2.0 infrastructure, refreshing it if needed. The exchange follows RFC 8693, OAuth 2.0 Token Exchange. The agent calls the tool with the exchanged token and never sees the provider token.

Consent is a separate surface from scopes. The broker's consent screen names a business capability, such as reading repositories, instead of presenting a raw scope string — Zalando's reasoning is that provider scope vocabularies are inconsistent and often too coarse for a person to evaluate. A user can reduce an agent's effective permissions on that screen when they are not comfortable granting everything it asks for. The broker also supports central tool approvals, so a tool call can be checked before an email is sent or a document shared, with a central audit trail.

What it does not solve

What to do with it

  1. Read the consent model before the code. The decision that matters is what permission sets your administrators define and what the consent screen shows users. A broker with permission sets as broad as the raw scopes underneath adds a vault and a log, not control.
  2. Try it against one service first. The repository's quickstart runs the broker locally with Docker Compose. Connect one third-party service, delegate one agent, and read the audit log for one tool call before wiring it into a fleet of agents and servers.
  3. Check the audit trail is actually reviewed. The broker writes structured JSON logs for compliance and monitoring, and records grant lifecycle changes. A log nobody reads is the same as no log — decide who reads it and when before the first delegation.
  4. Do not treat it as a substitute for scoping the agent. Keep the agent's own permissions narrow. The broker limits what a token can do; it cannot limit what an agent asks for, and a grant that covers everything the user can do is a grant the broker will faithfully honour.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Audit apps that installed agent plugins for you · Check your Codex and Claude Code config for relay tampering · Lock down Claude: IP allowlist and sign-out