Hi all,
@Stll0 suggested we bring this here from ns8-hermes-agent#40 to hear more opinions, so here it is.
The idea
Many of you have built Hermes agents on NethServer 8: instructions, skills, memory, the tools and the messaging gateways they use. We are building an agent application that runs as an NS8 module on the same cluster, with a web chat, workflows and voice agents that answer the phone through NethVoice.
We don’t want to replace Hermes. We want people who already run Hermes agents to use them from other apps too, with everything they built, instead of starting over. Everything stays on your NethServer, inside its internal network: nothing exposed to the internet, no new public route.
Two directions
1. Hermes uses tools from another app. Hermes is already an MCP client, and an agent pod can reach services on the node loopback. So another module can offer Hermes tools such as “call this number with a voice agent and tell me how it went”, or “start this workflow”. This already works today, with no change to the Hermes module.
Example: alice writes to her Hermes on Telegram: “call the three customers booked for tomorrow and confirm”. Hermes uses the voice tool, and tells her “two confirmed, one wants to move to Friday”.
2. Another app talks to your Hermes agent. You open a web chat, or a workflow reaches a step, and your own Hermes agent does the work as you configured it. Or someone phones the company, the voice agent asks the company’s Hermes agent, and answers the caller. This needs a change in the Hermes module, because today the agent’s API is reachable only inside its own pod.
What the Hermes module must keep guaranteeing
@Stll0 listed these requirements in the issue, and they stay:
- each agent can reach other apps on the same node;
- nobody reaches an agent’s interface without authentication;
- each user reaches only their own agent;
- agents can’t reach each other;
- no fork or patch of Hermes.
He would also like to simplify the current design (the socket relay and the custom auth proxy), using ns8-idp and Hermes’ native OIDC sign-in.
The open questions
- Sign-in. From the Hermes docs, its OIDC provider covers people on the dashboard, while the agent’s API still uses one key per agent. Is a key per agent, given only to modules the admin authorized at install time, and off by default per agent, the right shape for app-to-app access? Or is there a better NS8 pattern?
- One agent, one user. With ns8-idp, how would you keep “each user reaches only their own agent”? With a group, or a client per agent?
- Always on, ephemeral or sleeping? Stll0 suggested starting a Hermes instance on demand, with the agent’s configuration, skills and memory, or a sleep mode that wakes agents when needed. For reuse to work, it has to really be the same agent: one memory, and conversations that continue from one start to the next. How long would a start or a wake-up take? A web chat can wait a few seconds, a phone call much less.
- The phone line. Would you trust a Hermes agent to answer questions from callers, who are strangers? We think it needs a dedicated agent with read-only tools and no terminal. What would you require?
- Your use cases. If you run Hermes today: what would you want to reach it from (web chat, workflows, phone), and what should it never be allowed to do from there?
Thanks! We are happy to write the code and the docs for whatever shape comes out of this discussion.
Michele Vitiello Bonaventura • CallistoAI