DEV Community

jacobgadek
jacobgadek

Posted on

How to Handle Authentication for AI Agents (Without Hardcoding Cookies)

#security #ai #opensource #python

How to Handle Authentication for AI Agents (Without Hardcoding Cookies)

If you're building AI agents that interact with websites, you've probably run into this problem: your agent needs to be logged in.

The common solution? Hardcode cookies in your code:

# What everyone does (don't do this)
cookies = {"session_id": "abc123", "auth_token": "xyz789"}
response = requests.get("https://example.com/dashboard", cookies=cookies)
Enter fullscreen mode Exit fullscreen mode

This works until it doesn't:

  • Sessions expire and your agent breaks
  • Cookies leak into git repos
  • No way to track which agent accessed what
  • One leaked cookie = compromised account

The Real Problem: Agents Don't Have Identities

When a human logs into a website, they have an identity. They authenticate, get a session, and the site knows who they are.

AI agents? They're identity ghosts. They borrow your cookies, pretend to be you, and hope nothing breaks.

This is fine when you have one agent running locally. But what happens when you have:

  • Multiple agents accessing different services
  • Agents running in production
  • A team where everyone has their own credentials

You need a real solution.

What Good Agent Auth Looks Like

After dealing with this problem repeatedly, I built AgentAuth — an open-source SDK that handles this properly.

Here's how it works:

1. Export Cookies From Your Browser

Instead of manually copying cookies from dev tools, use a Chrome extension:

  • Go to the site (logged in)
  • Click the extension
  • Click "Export Session Cookies"
  • Done

2. Store Them Encrypted

Sessions are stored in an encrypted vault on your machine — not in your code, not in environment variables.

agent-auth init  # Set up the vault
agent-auth add github.com  # Paste the exported cookies
Enter fullscreen mode Exit fullscreen mode

3. Give Each Agent an Identity

Every agent gets a cryptographic identity (Ed25519 keys) and scoped permissions:

agent-auth create-agent github-bot --scopes github.com
Enter fullscreen mode Exit fullscreen mode

This agent can ONLY access github.com. It can't touch your LinkedIn or Gmail sessions.

4. Retrieve Sessions Securely 

from agent_auth import Agent, AgentAuthClient

agent = Agent.load("github-bot")
client = AgentAuthClient(agent)
cookies = client.get_session("github.com")

# Now use the cookies
response = requests.get("https://github.com/notifications", cookies=cookies)
Enter fullscreen mode Exit fullscreen mode

5. Everything is Logged

Every session access is logged. You know exactly which agent accessed which domain and when.

Why This Matters

Right now, there are thousands of AI agents running with hardcoded cookies. As agents become more common, this becomes a massive security hole.

Think about it:

  • Agents accessing company data
  • Agents running automations on your behalf
  • Agents with access to sensitive accounts

Do you really want all of that authenticated with copy-pasted cookies in a .env file?

Try It

AgentAuth is open source:

pip install agent-auth
Enter fullscreen mode Exit fullscreen mode

GitHub: https://github.com/jacobgadek/agent-auth

It takes 5 minutes to set up:

  1. agent-auth init
  2. Export cookies with the Chrome extension
  3. agent-auth add <domain>
  4. Use AgentAuthClient in your code

If you're building agents, give it a try. I'd love feedback.

Building something with AI agents? I'm working on making agent authentication a solved problem. Follow along on Twitter or check out the GitHub repo.

Top comments (0)