Skip to main content
Hermes Agent is Nous Research’s open-source autonomous AI agent. Unlike a session-based chat tool, it runs as a persistent daemon: it keeps memory across sessions, runs scheduled tasks, and writes its own reusable skills from experience. You drive it from a web dashboard at your *.suji.fr address. Hermes is an orchestrator — it calls out to an LLM provider you choose (Anthropic, OpenAI, or OpenRouter) rather than running model weights locally, so it runs fine on a normal CPU VM. You bring your own model key. This guide covers configuring Hermes on Suji. Check the version and release notes shown in the marketplace before installing or upgrading. Hermes is maintained upstream; Suji provides the marketplace packaging.
The Hermes workspace can approve and run shell commands. The current template requires a dashboard username and password and enables authenticated dashboard access. Choose a strong password and restrict access: the agent can execute commands within its configured environment. A login is not a guarantee that every tool or integration is safe.

​
Before you install

You need an API key for one model provider: Have the key ready for the install form.

​
Install

Dashboard → Apps → Hermes Agent → Install: Click Deploy. When the install is running, the dashboard is live at https://hermes-<random>.suji.fr over HTTPS through the tunnel.

​
First connection

  1. Open the install URL (https://hermes-<random>.suji.fr).
  2. You’re taken to a Hermes sign-in page — enter the dashboard username and password you set at install.
  3. You land in the Hermes workspace. Start a chat, set up cron tasks, or connect a messaging channel.
The template supplies your provider key to Hermes. Test a simple conversation and check provider access, credit and model availability if it fails.

​
What you can do with it

  • Persistent memory — Hermes remembers context across sessions, unlike a stateless chat.
  • Scheduled tasks — give it recurring jobs (cron) that run unattended.
  • Self-written skills — it turns successful workflows into reusable skills.
  • Messaging channels — connect Telegram, Discord, and other platforms so you can talk to it outside the dashboard. Configure these from the workspace after install.

​
Switching or adding a model provider

For a fresh setup, the install form supplies the selected provider key. In the current Hermes release, a provider key saved inside the Hermes dashboard takes precedence over that environment value. If you have saved a key there, change it in Hermes. Otherwise update the appropriate provider variables in Suji’s Variables tab and click Apply changes. Verify the selected model with a test conversation.

​
Day-to-day management

Hermes ships frequently and upgrades can include breaking changes, so its upgrade policy is breaking-changes-flagged — review its release notes and back up before choosing Upgrade. Suji’s confirmation does not guarantee migration compatibility.

​
Troubleshooting

Use the Dashboard username and password you set at install (username defaults to admin). Forgot the password? Update DASHBOARD_PASSWORD in Variables and click Apply changes.
Check the API key for your chosen provider is valid and has credit, then look at the Logs tab. Update the key where it is stored: in Hermes if saved there, or in Suji’s Variables otherwise.
Check both VM memory and the 4 GiB container limit. A larger VM helps host pressure but does not automatically raise the app’s container limit. Review the tools running and their memory needs.
The dashboard isn’t up through the tunnel yet — usually still starting (wait ~30–60 s). If it persists, the Logs tab shows why the container isn’t running.

​
Where things live

Uninstalling deletes the volume by default — that’s all the agent’s memory and skills. Choose keep data during uninstall to preserve it.
  • The template declares 8 GiB minimum/recommended host memory and limits the Hermes container to 4 GiB. Medium (8 GB) is the corresponding starting size. These are template estimates; monitor the actual tools and concurrent tasks you run.
  • The agent itself is light (it offloads reasoning to your LLM provider) — sizing is about the tools it runs, not the model.

​
Reporting issues