WorkOS unveils auth.md, an open protocol for AI agent sign-up

Most web applications offer no structured way for an AI agent to register and gain access. WorkOS has proposed a fix: auth.md, an open, OAuth-based protocol that simplifies agent integration without manual form-filling.

The problem auth.md addresses

Today, a developer who wants to connect an AI agent to an application has to fill in a registration form by hand, specify permissions and obtain an API key. The agent cannot do this automatically: there is no standard way to find out which flows a service supports, which permissions it requires and how to obtain credentials tied to a real user.

This becomes a bottleneck at scale. Every new agent needs manual configuration, and developers must write custom integrations for each application and keep the documentation up to date.

auth.md changes the approach. It is a plain Markdown file that an application publishes on its own domain (for example, example.com/auth.md). The file contains everything an agent needs to register on its own:

• Supported methods (OAuth 2.0, API keys, mTLS and others)
• Required scopes and permissions for different use cases
• Endpoints for obtaining credentials and long-lived tokens
• Code examples and recommendations for a quick start
• Information on rate limits, quotas and token lifetimes

How it works in practice

When an AI agent needs to integrate with an application, it goes to the application's domain, finds auth.md and reads the instructions. It then automatically starts the appropriate registration flow, requests the necessary permissions and receives a token or API key tied to a real user. A human does not have to fill in anything by hand: the whole process happens programmatically.

For example, if an agent connects to an email service, it can request exactly the permissions it needs (reading messages, sending replies) without access to other functions (deleting messages, changing account settings). Each permission is stated explicitly in auth.md, so the user sees what the agent is requesting and can approve or decline.

WorkOS chose Markdown over JSON or XML because the format is easy for both developers and AI agents to read and edit. A Markdown file is also easy to version in git, track changes to, annotate with comments and update without redeploying the application.

What this means for the ecosystem

auth.md is a step toward standardizing how AI agents integrate with the web. If the protocol gains enough support among major services, AI agents will be able to register on their own with hundreds of applications, without developers intervening manually. That would significantly speed up the automation of routine tasks.

However, new security questions arise: how do you prevent unauthorized access through agents? How do you manage permissions when registration is fully automatic? WorkOS expects that open standardization will make it easier to discuss these questions in the community and help develop best practices for secure automatic agent registration.