OpenClaw, Hermes Agent, or NemoClaw: Which to Self-Host
This hub covers three self-hosted AI assistant projects: OpenClaw (gateway-first, messaging integrations), Hermes Agent (agent-first, persistent cross-session memory plus a learning loop that builds reusable skill documents), and NemoClaw (NVIDIA's sandbox runtime, which wraps OpenClaw or Hermes rather than competing with either). All three run on Liquid Web infrastructure.
Stats on this page (GitHub stars, release versions, channel counts, RAM figures) were re-verified against their sources on 2026-09-10. Numbers that move fast are dated rather than stated as flat facts.
Two events in early 2026 changed the OpenClaw security landscape. OpenClaw's creator joined OpenAI in February 2026, raising enterprise questions about data sovereignty and long-term independence. In the same month, Koi Security reported that 341 of the 2,857 ClawHub skills it audited were malicious (the "ClawHavoc" campaign), and SecurityScorecard's STRIKE team reported that the first 24 hours of scanning found more than 40,000 internet-facing OpenClaw instances, 35.4% of observed deployments flagged as vulnerable to remote code execution, a count it says has kept growing since (SecurityScorecard STRIKE, published 2026-02-11). NemoClaw (announced at GTC 2026 on March 16) is NVIDIA's direct response: process-level sandboxing, network egress policies, and audit logging that contain any compromised skill or misconfigured agent. If you are running OpenClaw for any sensitive workload, read What Is NemoClaw before installing.
Which assistant fits your use case?
| Use Case | Recommended | Why |
|---|---|---|
| Personal productivity + messaging integrations | Either | Both ship a full messaging gateway. OpenClaw documents 31 channels (docs.openclaw.ai/channels, checked 2026-09-10); Hermes runs Telegram, Discord, Slack, WhatsApp, Signal and Email from one process. Decide on operating cost, not coverage |
| Repeated complex workflows | Hermes Agent | Persistent cross-session memory plus a learning loop, so the second run is faster than the first |
| Healthcare or enterprise compliance | NemoClaw, wrapping either | Process-level sandboxing, audit logs, policy enforcement. NVIDIA's README lists OpenClaw (default), Hermes, and LangChain Deep Agents Code as supported runtimes |
| Local GPU inference (70B models) | Hermes + Ollama or OpenClaw + Ollama | Both support Ollama; Hermes's learning loop compounds on local inference |
| Fewest moving parts to operate | Hermes Agent | One container, all state in SQLite, no database or vector store to run. OpenClaw's production stack adds Redis and Caddy alongside the gateway |
Healthcare or other regulated workloads need more than the table row above shows: HIPAA compliance means NemoClaw's sandboxing plus a signed Business Associate Agreement (BAA) with your host. See the HIPAA FAQ below for what that involves.
Not sure? Read the full comparison.
OpenClaw
OpenClaw (github.com/openclaw/openclaw) is a TypeScript AI assistant with ~389k GitHub stars (389,314, verified 2026-09-10) and an MIT license. It runs on your own machine or server, connects to just over 30 documented messaging channels (docs.openclaw.ai/channels, checked 2026-09-10), and supports local GPU inference and cloud API backends.
- What Is OpenClaw?: features, integrations, system requirements
- Self-Host OpenClaw on Liquid Web VPS: step-by-step Docker setup
- Docker Compose Stack: Redis, Caddy, backups, measured RAM footprint
- GPU Inference on Liquid Web: tier selection and Ollama/vLLM config
- HIPAA-Compliant Deployment: PHI guardrails and Liquid Web HIPAA setup
- Skill Development Guide: ClawHub SKILL.md format, test harness, publishing
- Monitoring & Observability: Prometheus, Grafana, alert runbook
NemoClaw
NemoClaw (github.com/NVIDIA/NemoClaw) is NVIDIA's sandbox runtime for agents. Announced at GTC 2026, it runs a supported agent inside an OpenShell sandbox with process-level isolation, network egress policies, and audit logging. NVIDIA's README lists OpenClaw as the default plus Hermes and LangChain Deep Agents Code (checked 2026-09-10), so choosing Hermes does not rule out NemoClaw. Apache 2.0, currently in Alpha.
- What Is NemoClaw?: architecture, inference modes, alpha caveats
- NemoClaw Policy Cookbook: Colang policy recipes for DLP, HIPAA, and skill gating
Hermes Agent
Hermes Agent (github.com/nousresearch/hermes-agent) is an open-source AI assistant from Nous Research (MIT, 243,861 GitHub stars verified 2026-09-10, latest release v0.21.1 published 2026-09-07). Unlike OpenClaw's gateway-first architecture, Hermes is agent-first: it plans, executes, reflects, and retries. It stores episodic and semantic memory in local SQLite with FTS5 full-text search, so it remembers your workflows across sessions with no external database to run. The memory layer is plugin-based and a different backend can be swapped in, but SQLite is the shipped default. After a multi-step workflow worth repeating, an error it worked past, or a user correction, Hermes writes a reusable skill document, so it gets faster the second time.
- What Is Hermes Agent?: architecture, memory types, system requirements
- Docker Compose Stack: one container, SQLite state in a single volume, ports 8642 and 9119, and the setup wizard you have to run first
- Learning Loop & Skills: guide and measure the learning loop; inspect skill documents
OpenClaw vs Hermes Agent vs alternatives
See vs Alternatives for a full comparison across OpenClaw, NemoClaw, Hermes Agent, Open WebUI, AutoGen, CrewAI, Dify, and more.
Resource footprint at a glance
| Tool | Idle RAM (remote LLM API) | Recommended Liquid Web tier |
|---|---|---|
| OpenClaw | ~180 MB gateway alone; ~235 MB idle / ~470 MB peak for the full Redis + Caddy + backup stack (measured 2026-05-01, local Docker 4 vCPU / 4 GB) | 4 GB Cloud VPS |
| OpenClaw + Ollama, local inference | Depends on the model; 7–13B fits the tier at right | L4 GPU server (24 GB VRAM) |
| Hermes Agent | No figure we can stand behind. Nous Research publishes no RAM minimum; its README says "Run it on a $5 VPS" (checked 2026-09-10) | 4 GB Cloud VPS |
| Hermes Agent + Ollama, local inference | Depends on the model; 70B at Q4 fits the tier at right | L40S GPU server (48 GB VRAM) |
The OpenClaw figures are our own measurements on local Docker (4 vCPU / 4 GB RAM), dated 2026-05-01; see the Docker Compose Stack guide for the per-service breakdown. Neither vendor publishes a RAM minimum: OpenClaw's docs give Node version requirements only (docs.openclaw.ai, checked 2026-09-10), and Nous Research's do the same for Hermes.
Correction (2026-09-10): this page previously listed Hermes at ~2 GB idle and recommended an 8 GB tier. That figure came from a measurement of a stack containing a PostgreSQL container Hermes has never required, so it measured the wrong thing. It is withdrawn rather than replaced, because we have not re-measured the current release. Budget the same 4 GB tier as OpenClaw and watch it under your own load.
For exact VRAM sizing by model and quantization, see the GPU inference guide. Prices subject to change, so verify at liquidweb.com/vps-hosting/managed-vps/ and liquidweb.com/gpu-hosting/.
OpenClaw is an open-source personal AI assistant (MIT license, TypeScript) with ~389k GitHub stars as of 2026-09-10. It runs on your own hardware or VPS, documents 31 messaging channels, supports local GPU inference via Ollama, and has a community skill marketplace, ClawHub, whose releases are moderated and whose manifests are signed (clawhub.ai, checked 2026-09-09).
NemoClaw is NVIDIA's open-source reference stack (Apache 2.0) that runs a supported agent in a sandboxed OpenShell environment with enterprise security controls. NVIDIA's README lists OpenClaw as the default plus Hermes and LangChain Deep Agents Code, checked 2026-09-10, so it is not OpenClaw-only. It was announced at GTC 2026 and is currently in Alpha.
Not by default. For HIPAA compliance you need NemoClaw's sandboxing, a Business Associate Agreement (BAA) with your hosting provider, encrypted storage, and audit logging. Liquid Web offers HIPAA-compliant dedicated servers and signs a BAA; pricing is quote-based.
Hermes Agent is an open-source AI assistant from Nous Research (MIT, ~244k GitHub stars as of 2026-09-10, latest release v0.21.1, published 2026-09-07). Unlike OpenClaw, it maintains persistent memory across sessions and runs a learning loop that generates reusable skill documents from successful task completions. That memory lives in local SQLite with FTS5 full-text search inside the container's data volume, not in PostgreSQL or a vector database, so there is nothing extra to provision.
No. Both run on a VPS using a cloud API backend like Claude or OpenAI. GPU is only needed for fully local inference via Ollama or vLLM, which eliminates cloud API costs and keeps all data on your hardware.