Seven end-to-end guides for deploying HermesShell in real scenarios. Each guide covers prerequisites, step-by-step setup, verification, and a NemoClaw comparison.
| Who you are | Stack | Guide |
|---|---|---|
| Researcher / writer | Docker + Telegram + arXiv digest | 01-researcher |
| Developer | Docker + VS Code ACP integration | 02-developer |
| Home automation | Docker + Home Assistant MCP + Telegram | 03-home-automation |
| Data analyst | Docker + Postgres MCP + anomaly alerts | 04-data-analyst |
| Small business | Docker + Slack support bot | 05-small-business |
| Privacy-regulated industry | OpenShell sandbox + strict policy | 06-privacy-regulated |
| Trader / quant | Docker + Qwen-7B + Telegram alerts | 07-trader |
NemoClaw (alpha, March 2026) is NVIDIA's reference stack for OpenClaw agents. It provides the OpenShell sandbox, 25+ tools, messaging gateways (Telegram, Discord, Slack, WhatsApp, and more via OpenClaw), and supports multiple inference providers (OpenAI, Anthropic, Google Gemini, NVIDIA NIM). Local model inference is broken on macOS (DNS bug, issue #260) — cloud APIs required on macOS.
| Use case | HermesShell | NemoClaw | Key difference |
|---|---|---|---|
| Researcher (cross-session memory + digest) | ✅ | NemoClaw: Telegram ✅, but no persistent MEMORY.md/USER.md across sessions | |
| Developer (VS Code ACP) | ✅ | NemoClaw: no ACP — OpenClaw has its own IDE integration | |
| Home automation (HA MCP + Telegram) | ✅ | NemoClaw: Telegram ✅, but no MCP server support (unconfirmed) | |
| Data analyst (Postgres MCP + alerts) | ✅ | NemoClaw: Slack/Telegram alerts ✅, but no Postgres MCP (unconfirmed), no persistent memory | |
| Small business (Slack bot) | ✅ | ✅ | Both support Slack natively — difference is local vs cloud inference |
| Privacy-regulated (air-gapped, local inference) | ✅ | NemoClaw: sandbox ✅, but local inference broken on macOS — cloud API required (data leaves network) | |
| Trader (local latency + Telegram) | ✅ | NemoClaw: Telegram ✅, but local inference broken on macOS — cloud API adds 200–500ms latency |
Summary: NemoClaw has more capabilities than previously documented — messaging gateways, voice, and multi-provider inference. The remaining HermesShell advantages are: persistent cross-session memory (MEMORY.md/USER.md), self-improving skills (DSPy + GEPA), MCP server support, and local inference that works on macOS. For privacy-sensitive and low-latency use cases, local inference is the decisive factor.
Each use case has a corresponding installable Hermes skill. Skills are instruction files that Hermes reads to execute recurring workflows (weekly digests, anomaly detection, market alerts).
# Install from the repo root
./skills/install.sh research-digest # Researcher
./skills/install.sh code-review # Developer
./skills/install.sh home-assistant # Home automation
./skills/install.sh anomaly-detection # Data analyst
./skills/install.sh slack-support # Small business
./skills/install.sh market-alerts # TraderFull documentation: skills/README.md
All Docker-mode guides share these baseline requirements:
- Docker Desktop (or Docker Engine + Compose) — install
- A GGUF model file in
models/— recommended:Qwen3-4B-Q4_K_M.gguf(~2.5 GB) - Git to clone the repo
git clone https://github.com/ppritcha/hermesshell
cd hermesshell
cp .env.example .env
# Drop your .gguf model into models/OpenShell-mode (guide 06) additionally requires:
- Linux host (Ubuntu 22.04+) or macOS Apple Silicon with Colima
- NVIDIA OpenShell runtime — install