Skip to main content

Run an agent you do not have to trust

Most AI agent demos give a model a terminal and hope for the best. That breaks down when the agent follows a bad prompt, pulls a risky dependency, or treats the whole internet as fair game.

This demo is different. With Hermes you build a friendly weather agent, pack it into AgentPaaS, and see real product security:

  • Isolated container, not your laptop user account
  • Default-deny network - only hosts you approve (weather + LLM)
  • Brokered secrets - API keys never sit in agent code or chat
  • Lineage and audit - proof the agent is secure and is being governed and audited

You finish with the same agent on your Mac and on AgentPaaS Cloud.

Follow the steps in order. Prefer pasting the Hermes lines as written.

What you need​

MacHermes + local AgentPaaS runtime
HermesInstall Hermes if needed
Trial inviteClaim email from AgentPaaS (get it here)
LLM keyPrefer OpenRouter and a cheap model

Step 1: Claim your trial​

Open the claim link from your email in a browser you will keep using. Set a password (Google optional). Use the same email the trial was issued for.

You are done when you can open cloud.agentpaas.ai and see the console.

Step 2: Install AgentPaaS in Hermes​

In Hermes, paste:

Install from https://github.com/AgentPaaS-ai/agentpaas

Hermes installs the plugin and local tooling. If tools are missing, /quit, reopen Hermes, and paste again.

Step 3: Add your LLM key (you paste once, privately)​

Ask Hermes:

Set up my publisher identity and OpenRouter credential for a weather agent. Coach me to add the secret in MY terminal only, then confirm labels only with secret list.

When Hermes asks you to run a command, do it in your Terminal window and paste the key at the prompt. Never put the key in chat.

Step 4: Build the weather agent​

In Hermes:

Build a weather agent that uses an LLM, and responds in a friendly demeanour

Keep egress tight: weather host + OpenRouter (or your LLM host) only. Everything else is denied by default.

You are ready when a local invoke returns a friendly weather answer.

Step 5: The big reveal - lineage and audits​

In Hermes:

Show me lineage and audits

Lineage is the signed build story of the agent artifact (who packed what version and digest).
Audit is the run log under policy, including egress_allowed and egress_denied.

If the agent tried a website that was not on the allow list, you should see a denial recorded by the gateway. That is the product working.

Optional teaching moment:

Demonstrate governance: remove the weather host from policy, repack, invoke and show the denial in the audit. Then add the host back, confirm with me, repack, and invoke successfully.

Step 6: Run it on AgentPaaS Cloud​

In Hermes:

Make it run in the AgentPaaS cloud

When login is needed, Hermes should not hang in a browser for you. In your Terminal:

agentpaas cloud login

Approve in the same browser you used to claim the trial, then tell Hermes to continue (push, deploy, bind secret, invoke).

Step 7: Look in the console​

Open cloud.agentpaas.ai and check Agents, Deployments, and Runs. Expand a row to see detail.

Optional schedule (ask Hermes in plain language):

Schedule this cloud deployment every 5 minutes, then show me how to disable it.

Or explore tabs yourself: Console tour.

What you proved​

  1. The agent never got the open internet - only approved hosts.
  2. Denials and allows are audited.
  3. Secrets stayed out of chat and agent source.
  4. Lineage ties the artifact to a signed build.
  5. The same governed agent runs locally and in cloud.

That is AgentPaaS: a secure place to run agents.

If something fails: Troubleshooting.