Architecture

The whole picture

HermesCo is a Next.js application (App Router) that hosts the agent loop, the Treasury, and the Command Center UI. The agent reasons with Hermes 4 405B, every spend is screened by NVIDIA Nemotron, money moves through Stripe, and the agent’s real work runs on a dedicated Fly Machine. State persists in Convex when configured, with an in-memory fallback otherwise.

                          Command Center  (/command)
                                  |
                                  v
   POST /api/hermesco/agent  -->  runTurn()           src/lib/hermesco/agent.ts
        |                            |
        |                            |  Hermes 4 405B via OpenRouter
        |                            |  drives tools through <tool_call> XML
        |                            v
        |                       executeTool()          src/lib/hermesco/tools.ts
        |                        /     |        \
        |          check_treasury  propose_spend   run_in_sandbox
        |                |             |                  |
        |                v             v                  v
        |          Treasury       screenSpend()      runForAgent()
        |       treasury.ts        safety.ts          sandbox.ts
        |                |             |                  |
        |                |   Nemotron + hard caps    Fly Machine (or Daytona)
        |                v                               fly.ts
        |             store.ts  ----> Convex or in-memory
        |
        +--> Stripe (Checkout, payment links, PaymentIntents)  stripe-skills.ts

Components

Brain

Hermes 4 405B reasons and emits <tool_call> blocks. The server parses them, runs the tool, feeds the result back, and loops up to a bounded number of steps.

Safety screen

Every spend passes a deterministic policy gate and an NVIDIA Nemotron classification before it can execute.

Money rail

The Treasury holds real capital and records a signed ledger. The agent’s Stripe skills (server-side wrappers over the official Stripe API) create offers, collect revenue, and run deposits.

Agent body

Each agent is a dedicated Fly Machine. The agent execs real commands on it, and it suspends to $0 when idle.

Request lifecycle of one directive

The Command Center posts the directive and conversation history to POST /api/hermesco/agent.
runTurn sends the system prompt, tool specs, and history to Hermes 4 405B through OpenRouter.
Hermes responds with prose and zero or more <tool_call> blocks containing JSON.
The server executes each tool. Money tools route through the Treasury and the safety screen.
Tool results are appended as <tool_response> and Hermes is called again, up to MAX_STEPS.
The loop ends when Hermes stops calling tools or a spend is awaiting human approval. The turn returns a structured AgentTurnResult (events, assistant text, awaiting-approval flag, and the live Treasury state).

Why a server-parsed tool protocol

Hermes 4 405B on OpenRouter does not expose native function calling, so HermesCo drives tools through Hermes’s own <tool_call> convention and parses them server-side. This keeps Hermes genuinely in control of the money and infrastructure tools, which is the authentic Nous approach. Nemotron rides underneath as the safety classifier rather than as a competing brain.

Source map

Area	File
Agent loop	`src/lib/hermesco/agent.ts`
Tools	`src/lib/hermesco/tools.ts`
Models and chat client (OpenRouter + NVIDIA)	`src/lib/hermesco/models.ts`
Safety screen	`src/lib/hermesco/safety.ts`
Treasury	`src/lib/hermesco/treasury.ts`
Stripe Skills	`src/lib/hermesco/stripe-skills.ts`
Agent body (Fly)	`src/lib/hermesco/fly.ts`
Sandbox router	`src/lib/hermesco/sandbox.ts`
Persistence	`src/lib/hermesco/store.ts`
Identity	`src/lib/hermesco/useIdentity.ts`
Domain types	`src/lib/hermesco/types.ts`
Command Center	`src/app/command/page.tsx`

Get Started

Concepts

Operations

About

The whole picture

Components

Brain

Safety screen

Money rail

Agent body

Request lifecycle of one directive

Why a server-parsed tool protocol

Source map

​The whole picture

​Components

Brain

Safety screen

Money rail

Agent body

​Request lifecycle of one directive

​Why a server-parsed tool protocol

​Source map

The whole picture

Components

Request lifecycle of one directive

Why a server-parsed tool protocol

Source map