Guide
Hermes Agent System Requirements
Hermes Agent runs comfortably on modest hardware for chat-only workloads, but the browser harness and large model contexts push memory and disk fast. Here's the minimum, the recommended, and where the bottlenecks actually are in 2026.
At a Glance
| Minimum | Recommended | |
|---|---|---|
| RAM | 1 GB | 2–4 GB |
| CPU | 1 vCPU | 2 vCPU |
| Disk | 2 GB | 10 GB+ (with browser harness) |
| Node.js | 20.x | 22.x LTS |
| OS | Linux / macOS / Windows (WSL2) | Linux x86_64 or arm64 |
| GPU | None | None (models are remote) |
| Network | Outbound 443 | Outbound 443 + inbound for webhooks |
RAM
A chat-only Hermes container holds steady around 300–600 MBresident. Adding the browser harness (Chromium for tool use) pushes peak resident usage to 1.2–1.8 GB, especially with multiple tabs open. Large models that send long context windows back through the agent (Claude 200k, Gemini 1M) push working set higher transiently.
For a single-user VPS, 2 GB is comfortable. For multi-user managed hosting, plan on ~1 GB per concurrent active container.
CPU
Hermes Agent is mostly I/O bound — it waits on model responses, network webhooks, and disk reads. 1 vCPU is enough for low-volume use; 2 vCPU helps when the browser harness is rendering pages or skills do CPU-heavy parsing (PDF, video).
Disk
The base install is small (under 500 MB), but the workspace at ~/.hermes grows with chat history, embeddings, and any files your agent touches. The browser harness pulls in Chromium (~400 MB) and per-session profile data. Plan on:
- 2 GB — minimum, chat-only, ephemeral workspace
- 10 GB — comfortable for browser harness + memory store
- 20 GB+ — if you keep long chat history and large skill data
Node.js Version
Hermes Agent requires Node 20 or newer. Node 22 LTS is recommended for new installs. Node 18 reached EOL and is not supported. nvm works well if your distro ships an older version.
OS Support
- Linux — first-class, x86_64 and arm64
- macOS — Intel and Apple Silicon, dev environments only
- Windows — via WSL2 (native Windows is best-effort)
- Docker — the official image works on any OS that runs Docker
GPU? No.
Hermes Agent does not run local model inference. Model calls go to remote providers (OpenRouter, Anthropic, OpenAI, etc.) over HTTPS. You do not need a GPU. If you want local inference, point Hermes at a local Ollama or LM Studio endpoint — the GPU lives on the inference server, not the Hermes box.
Network
- Outbound 443 to your model provider (required)
- Inbound 80/443 if you accept channel webhooks (Telegram, Slack, etc.)
- Outbound to ClawHub / skill registries if installing skills at runtime
Sizing Cheat Sheet
- Personal bot, chat-only: 1 vCPU / 1 GB RAM / 5 GB disk — any $4–5/mo VPS
- Personal bot + browser harness: 1 vCPU / 2 GB RAM / 10 GB disk
- Team bot, 5–10 concurrent users: 2 vCPU / 4 GB RAM / 20 GB disk
- Production, 50+ concurrent: 4 vCPU / 8 GB RAM / dedicated node
Skip the Sizing
OpenClaw Launch hosts Hermes Agent on right-sized containers with the browser harness, memory store, and channel webhooks pre-wired. Plans from $3/mo. See also the Hermes VPS guide for self-hosting recommendations.