Hermes on NemoClaw

Tool-using AI agents as a product, running fully on-premise on Dell workstations — policy-bounded agent sessions, sandboxed tools, and private inference on a single deskside system.

Summary

Hermes Agent, built by Nous Research, is the tool-using agent harness this solution deploys: a policy-bounded autonomous agent that reasons with a locally-served LLM (vLLM, OpenAI-compatible), calls external tools through the Model Context Protocol (MCP), and keeps skills and memory across sessions. It runs on the NemoClaw runtime — NemoClaw supplies policy, lifecycle, and the gateway; the OpenShell sandbox supplies filesystem landlock and L7 egress boundaries. Everything runs on one Dell deskside system; no cloud dependency for the model path.

This solution is the deployment path for Hermes on Dell S / M / L tiers: validated sizing, one-command serving recipes, and an IT-Admin fleet guide. The same local endpoint (http://localhost:8000/v1) also serves the Agentic Coding solution on the same box. For the base runtime with its default OpenClaw agent, see OpenClaw on NemoClaw.

What this is not: an IDE autocomplete solution (that is Agentic Coding), nor an unrestricted internet agent — policy and egress boundaries are the point.

Personas

Terms: Glossary.

Reference NVIDIA playbooks

Use cases

Use cases this solution enables (NVIDIA NemoClaw / Hermes positioning, mapped onto Dell tiers with local vLLM):

Deskside pitch: you feel the agent on your own GPU, not a shared cloud tenant.

Architecture

The agent loop, end to end:

        User / channel / CLI
                │
                ▼
        Hermes agent   ── harness: reasoning loop,
                │          tool-calling, skills, memory
                ▼
        NemoClaw runtime ── policy, lifecycle, gateway
                │
    ┌───────────┴────────────────┐
    ▼                            ▼
 OpenShell sandbox         Local LLM — vLLM :8000 (/v1)
 (landlock + L7 egress)         │
    │                           ▼
    ▼                    Nemotron 3.5 Lightning (NVFP4)
 MCP tool servers
 (files, git, tickets, APIs)

        Dell deskside workstation (S / M / L tier)
_Commands in this guide run from the repository root (`git clone` this repo and `cd` into it)._

Two-plane deployment

Plane 1 is reusable across solutions — the same http://localhost:8000/v1 endpoint serves Hermes, coding assistants, and Agentic Video Search & Summarization on one box.

Key components

S / M / L sizing

Agent workloads have variable, stateful resource consumption — these are conservative baselines from measured tool-using agent ladders. One Hermes task ≈ several sequential LLM calls plus tool executions, so serving concurrency ≠ agent concurrency; size from the agent column, not raw tokens/sec. Sizing framework: S/M/L Sizing.

Model recommendation

• Default: Nemotron family, scaled per tier. nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 — tool-calling capable MoE (30 B total / ~3 B active), NVFP4; runs on every tier, so a mixed fleet shares one default. M / L target pick is Nemotron 3.5 Super VL (recipes/{precision,gb300}/vllm/); Lightning is the S-tier default.
• Tier alternatives (same OpenAI-compatible API, tool-calling verified):
• S — nvidia/Qwen3.8-27B-NVFP4 (+ MTP) · recipes/gb10/vllm/docker-qwen3.8-27b-nvfp4.sh
• M — nvidia/Qwen3.8-Flash-Next-NVFP4 (MTP ×3, engram CPU offload) · recipes/precision/vllm/docker-qwen3.8-flash-next-nvfp4.sh
• L — RedHatAI/GLM-5.3-Flash-NVFP4 (320 B/18 B-active MoE, 262 K ctx) · recipes/gb300/vllm/docker-glm-5.3-flash-nvfp4.sh

Agentic tool-calling quality drives model choice — validated function-calling is a hard requirement for Hermes. The shipping recipes bake the default model's flags (--tool-call-parser qwen3_coder + --reasoning-parser nemotron_v3); parser semantics: vLLM tool-calling docs. Validate any alternate model with the smoke test in Developer Deploy Guide before adopting.

Inference stack

vLLM (preferred). OpenAI-compatible, production-grade. Hermes connects to http://localhost:8000/v1 — the shared endpoint across this repository's solutions; models are shared across workloads on one box. Alternative stacks (ollama, llama.cpp, NIM): Inference Stack & Models.

Deployment recipes

Two-step quick start (for a fleet, scripts/landscape/deploy-solution.sh runs both steps):

# 1. Serve the local LLM (shipping default = Nemotron 3.5 Lightning)
recipes/gb10/vllm/docker-nemotron-3.5-lightning-nvfp4.sh --port 8000      # Dell Pro Precision with GB10
# recipes/precision/vllm/docker-nemotron-3.5-lightning-nvfp4.sh --port 8000  # Dell Pro Precision 9 T4 (RTX PRO 6000)
# recipes/gb300/vllm/docker-nemotron-3.5-lightning-nvfp4.sh --port 8000       # Dell Pro Precision with GB300

curl -s http://localhost:8000/v1/models

# 2. NemoClaw installer + Hermes onboarding
curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash
nemoclaw onboard             # choose Hermes / nemohermes when prompted
nemohermes agents list       # expect: openclaw | hermes | ...

Full per-plane steps: Developer Deploy Guide (developers) · IT-Admin Guide (fleet). Recipe catalog: Deployment Recipes.

Benchmarks & performance (validation)

Agent capacity is stated in concurrent agents and per-turn wall-time, not raw tokens/sec. Same tool-using workload, same MCP tool set, same shipping-default model (Nemotron 3.5 Lightning NVFP4) on every line — only the Dell system changes. All tiers pass at concurrency 1 / 2 / 4 with 100 % task success.

Serving-layer context (not the agent envelope): the default model on a single RTX PRO 6000 sustains 32 – 64 concurrent serving requests at p99 TTFT ≤ 264 ms — for sizing a multi-agent gateway fanning into one Plane 1 endpoint, not for "how many analysts share this box?" (use the agent table).

Management environment

• Identity: Hermes inherits the host's identity by default; on multi-user M / L deployments, front vLLM with an AD/LDAP-backed reverse proxy.
• Monitoring: vLLM /metrics scraped by Prometheus, Grafana (:3000); NemoClaw exposes agent telemetry (task success, agent turns, tool-call counts).
• Agent policies: standard / strict / permissive tiers ship with NemoClaw — ship standard in production; apply at sandbox creation (IT-Admin guide).
• Fleet: containerized deploy (Docker / Podman); scale by adding agent instances behind a shared gateway on M / L tiers.

Dependencies & Risk factors

• Dependencies: NemoClaw installer + Hermes agent from NVIDIA; vLLM on GB10 / RTX PRO / GB300; NVFP4 model availability; validated tool-calling in the served model (hard requirement).
• Risk factors: agent resource use is variable and stateful — envelopes are conservative; NVFP4 serving needs the recipe-pinned backend/parser flags; model recommendations evolve with the Nemotron family.

Related

• Developer Deploy Guide — developer quick-start · IT-Admin Guide — IT-admin fleet guide.
• OpenClaw on NemoClaw — the same runtime with the default OpenClaw agent, no Hermes.
• Agentic Coding — coding-assistant companion sharing the same vLLM serving layer.
• Personas — persona definitions.
• NVIDIA entry point: NemoClaw playbook.