Skip to main content

​
One agent, not a toggle

HermesCo runs a single operator. There is no model picker and no “Hermes or Nemotron” switch. Hermes is the brain; Nemotron is a safety screen that is always on underneath it.

​
Hermes decides

Hermes 4 405B is the operator. It receives a system prompt, the available tool specs, and the conversation history, then reasons and emits tool calls in Hermes’s native format:
The server parses each <tool_call>, executes the tool, and appends the result as a <tool_response> for the next step. The loop is bounded so a turn always terminates. Hermes stops calling tools when the task is done or when a spend is awaiting a human decision. The model is defined in src/lib/hermesco/models.ts:

​
Nemotron screens

Every proposed spend is classified by NVIDIA Nemotron before it can move money. This is the NemoClaw layer. It is not a different agent you choose; it is a second set of eyes that runs on every spend.
Nemotron’s judgement is layered with deterministic checks so the safety outcome never depends on a model being available. See Safety screen for the exact order of evaluation.

​
Stripe settles

When money actually moves, it moves through Stripe: real Checkout sessions for deposits, real payment links for earning, and a real ledger entry for every spend. There is no simulated fallback and no canned test card. If Stripe is not connected, the money tools say so rather than faking a result. See Treasury.

​
Why this unifies the three sponsors

The hackathon brings together Nous Research (Hermes), NVIDIA (Nemotron, NemoClaw, agent skills), and Stripe (payment skills). Rather than hiding them behind a dropdown, HermesCo composes them into a single operating loop: the Nous brain decides, the NVIDIA screen verifies, and the Stripe rail settles, all on real infrastructure.