TechLatest

Posted on Jun 10 • Originally published at Medium on Jun 9

OpenClaw or Hermes? Choosing the Right AI Agent Stack in 2026

#agents #hermes #openclawaiagent #hermesagent

The AI model race is slowing down. The agent runtime race is just getting started.

In 2025, everyone compared Claude, GPT, Gemini, and Qwen. In 2026, the conversation has shifted. The real question is no longer which model you use, but which system orchestrates that model.

For self-hosted agents, two projects stand out: OpenClaw and Hermes Agent.

Both can connect to Telegram, Discord, Slack, WhatsApp, local tools, and cloud models. Both support skills. Both can automate tasks and execute workflows.

Yet after spending time with both systems, I came away with a simple conclusion:

OpenClaw is a better control plane. Hermes is a better self-improving runtime.

The choice depends entirely on what you expect your agent to become.

Repos: NousResearch/hermes-agent · openclaw/openclaw

Part 1 — What problem do they solve?

At first glance, OpenClaw and Hermes look similar.

You connect a model.

You give it tools.

You chat with it through Telegram, Discord, WhatsApp, or the terminal.

But their philosophies diverge quickly.

OpenClaw treats agents as members of a larger system.

Hermes treats agents as individuals that learn and improve over time.

That difference influences everything else.

| Category | OpenClaw | Hermes |
| ------------------ | ------------------------------------------------------- | ---------------------------------------------------- |
| **Core Idea** | Agent control plane | Self-improving runtime |
| **Primary Focus** | Channels, routing, and orchestration | Learning, memory, and automation |
| **Ideal User** | Operators, builders, and teams managing multiple agents | Researchers, automation enthusiasts, and power users |
| **Long-Term Goal** | Manage and coordinate many agents | Continuously improve a single agent over time |

Both projects answer: “How do I talk to an AI agent from Telegram/WhatsApp/Discord and have it use tools on my machine?”

They diverge on what happens after the first week :

| | OpenClaw | Hermes |
|---|----------|--------|
| **Product feel** | Polished personal assistant — gateway, channels, dashboard | Research-grade agent platform — tools, memory, evolution |
| **Skills** | You install or write `SKILL.md`; ClawHub registry | Agent can **author** skills; Curator maintains quality |
| **Stack** | Node.js, TypeScript, npm global | Python CLI, bash installer |
| **Sweet spot** | "Message my assistant anywhere" | "My assistant gets better at my workflows over time" |

Neither is a hosted SaaS. You run the gateway on your laptop, homelab, or VPS.

Part 2 — Architecture side by side

OpenClaw

Gateway = single control plane (default http://127.0.0.1:18789/))
Workspace = ~/.openclaw/workspace with AGENTS.md, SOUL.md, TOOLS.md
Skills = ~/.openclaw/workspace/skills//SKILL.md
Daemon = launchd/systemd user service after openclaw onboard --install-daemon

Docs: Architecture · Gateway

Hermes

CLI + TUI = hermes, hermes --tui
Gateway = hermes gateway for messaging platforms
Skills = procedural memory in ~/.hermes/skills/
Curator (v0.12+) = periodic grading/pruning of learned skills

Docs: Hermes user guide

Shared pattern

Both normalize inbound chat JSON → agent message → tool/skill execution → outbound reply. Both use Markdown skills as the extension point for custom workflows.

Architecture Verdict

Choose OpenClaw when:

You need multiple agents
You need channel separation
You need orchestration

Choose Hermes when:

You want a single powerful assistant
You care about automation
You value simplicity

Winner: OpenClaw

Deploy on OpenClaw VM

Want to skip infrastructure setup?

We provide pre-configured OpenClaw VM images on [AWS](https://aws.amazon.com/marketplace/pp/prodview-y7ck4mk5qmrdk?utm_campaign=openclaw-vm&utm_source=techlatest-website&utm_medium=support-page\), Azure, and Google Cloud Platform (GCP). Each deployment comes with OpenClaw, Ollama, and all required dependencies pre-installed, allowing you to launch a production-ready AI agent environment in minutes.

Available with both CPU and GPU configurations for development, testing, and production workloads.

Skills: Static Catalog vs Living Knowledge

This is where Hermes becomes interesting.

OpenClaw uses a traditional skill ecosystem.

You install skills.

You update skills.

You manage skills.

The model stays mostly separate from the skill lifecycle.

Hermes takes a different approach.

Repeated workflows can become reusable skills.

Instead of treating skills as software packages, Hermes treats them as procedural memory.

Over time, the agent begins to recognize recurring patterns and formalize them.

This fundamentally changes the relationship between user and system.

With OpenClaw, you manage skills.

With Hermes, you train skills.

Skills Verdict

If you want predictability:

OpenClaw

If you want adaptation:

Hermes

Winner: Hermes

Memory: Rich Context vs Focused Context

Memory is often marketed as a feature.

In reality, memory is usually a tradeoff.

OpenClaw maintains richer context across workflows and channels.

That can be incredibly useful.

It can also create noise.

As systems grow, context retrieval becomes harder to manage.

Hermes intentionally keeps memory lean.

Instead of aggressively pulling context into every task, it retrieves information progressively.

The result is a system that often feels more focused.

OpenClaw remembers more.

Hermes remembers more selectively.

Memory Verdict

For long-running agent ecosystems:

OpenClaw

For daily workflows and repeated tasks:

Hermes

Winner: Hermes

User Experience and Control

This was one of the most surprising differences.

OpenClaw generally feels mature and stable.

Once configured, it stays out of the way.

Hermes feels more transparent.

Tool execution is easier to inspect.

Context usage is easier to understand.

Interrupting workflows feels more natural.

If you enjoy seeing what your agent is doing, Hermes provides a clearer window into the system.

If you simply want the system to work, OpenClaw’s maturity is reassuring.

UX Verdict

Transparency: Hermes

Stability: OpenClaw

Overall Winner: Hermes

Part 3 — Prerequisites

| Requirement | OpenClaw | Hermes |
|-------------|----------|--------|
| OS | macOS, Linux, Windows (WSL2) | macOS, Linux, WSL |
| Runtime | Node **22.19+** or **24** | Python (installer handles deps) |
| API key or local model | Yes | Yes |
| Disk | ~500MB+ for Node + workspace | ~1GB+ depending on browser tools |

Check versions:

node -v # v22.19+ or v24 for OpenClaw
which hermes # after Hermes install
which openclaw # after OpenClaw install

Part 4 — Install OpenClaw

npm install -g openclaw@latest
openclaw onboard --install-daemon

The onboarding wizard configures:

Gateway bind address and auth
LLM provider (or Ollama for local models)
At least one channel (Telegram is the fastest smoke test)
Workspace path and bundled skills

Verify:

openclaw doctor
openclaw status
# Dashboard (if gateway running):
# http://127.0.0.1:18789/

Local model (optional): follow the OpenClaw + Gemma + RAG tutorial to point OpenClaw at gemma4:e2b via Ollama.

OpenClaw skills smoke test

openclaw skills list
openclaw skills install <skill-from-clawhub> # example — see clawhub.ai

Skills load from (highest priority first):

/skills/
Project /.agents/skills
~/.agents/skills
~/.openclaw/skills
Bundled skills

See Skills docs.

Part 5 — Install Hermes

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
source ~/.zshrc # or ~/.bashrc
hermes setup --portal

hermes setup --portal is the fastest path to a working cloud model + tool gateway. For local-only, use hermes model and configure Ollama per Hermes docs.

Verify:

hermes doctor
hermes --tui

First TUI prompts to try:

“List tools you have access to”
“List skills in ~/.hermes/skills”
“What is the Curator and when does it run?”

Full Hermes depth: Awesome Hermes Agent tutorial.

Hermes gateway smoke test

hermes gateway

Configure channel tokens via hermes setup or config files. Run hermes doctor after any gateway change. Keep DM pairing/allowlists enabled until you trust exposure.

Part 6 — Feature comparison (hands-on)

Use the same three prompts on both systems and compare behavior.

| Test prompt | What to observe |
|-------------|-----------------|
| *"What skills do you have?"* | OpenClaw lists workspace/ClawHub skills; Hermes lists `~/.hermes/skills` + may mention learned skills |
| *"Run a shell command: uname -a"* | Tool permission / sandbox behavior |
| *"Remember that my project codename is NEPTUNE"* | Memory persistence on next session |

Record results in a simple table:

| Test | OpenClaw | Hermes |
|------|----------|--------|
| Skill list | | |
| Shell tool | | |
| Memory | | |

Full static matrix: feature matrix.

Part 7 — Skills: same format, different lifecycle

OpenClaw skill anatomy

~/.openclaw/workspace/skills/my-skill/
├── SKILL.md # YAML frontmatter + instructions
└── scripts/ # optional Python/shell helpers

Install from ClawHub:

openclaw skills install <skill-id>
openclaw skills verify <skill-id> # trust envelope when available

Operator maintains skills — update via openclaw skills update or ClawHub sync.

Hermes skill anatomy

~/.hermes/skills/my-skill/
└── SKILL.md

Invoke explicitly: /skill my-skill or let the agent auto-select.

Learning loop: after repeated workflows, Hermes can draft new SKILL.md files from session traces. Curator (v0.12+) reviews and prunes them on a ~7-day cycle so quality does not drift.

Porting a skill between stacks

Copy the skill directory to the other runtime’s skills path.
Adjust tool names in SKILL.md (OpenClaw vs Hermes tool schemas differ).
Update any script paths (~/.openclaw ↔ ~/.hermes).
Restart gateway / start a new session.

Example: our agentic-rag skill targets OpenClaw — a Hermes port would call the same LitServe RAG API with Hermes shell tool syntax.

Part 8 — Channels & gateway

| Concern | OpenClaw | Hermes |
|---------|----------|--------|
| Start daemon | Installed by onboard | `hermes gateway` (or systemd per your setup) |
| Multi-channel | One gateway, many channels | One gateway, 18+ platforms |
| Config | `openclaw.json` + wizard | Hermes config under `~/.hermes/` |
| Chat commands | `/status`, `/new`, `/restart`, … | Hermes TUI + channel-specific |

Recommendation: enable one channel (Telegram) on both for comparison, then expand. Running both gateways on the same bot token will conflict — use separate bots or run one at a time.

Part 9 — Models: cloud vs local

OpenClaw + Ollama (this repo’s pattern)

ollama pull gemma4:e2b
# Configure in openclaw.json — see openclaw-gemma-rag/config/
openclaw gateway restart

Hermes + local model

Configure via hermes model or provider section in Hermes docs. Cloud APIs remain the path of least resistance for tool-heavy tasks on modest hardware.

| Workload | Suggestion |
|----------|------------|
| Phone assistant, mostly chat | Cloud model on either stack |
| Private docs, RAG, homelab | OpenClaw + [Gemma RAG guide](https://ayush7614.github.io/agentic-ai-ecosystem/guides/openclaw-gemma-rag/) |
| Heavy browser automation | Hermes with sandbox backend (Modal/Daytona) or skip browser on small VPS |

Part 10 — Memory & self-improvement

| | OpenClaw | Hermes |
|---|----------|--------|
| **Session history** | Session tools (`sessions_history`, etc.) | Built-in session + TUI history |
| **Long-term memory** | Workspace files + operator-managed | Memory layer + ecosystem plugins (honcho, hindsight, plur) |
| **Automatic skill growth** | No | **Yes** — core differentiator |
| **Quality control** | Manual review, `openclaw skills verify` | **Curator** automated

Choose Hermes when you want the agent to accumulate procedural memory. Choose OpenClaw when you want predictable, curator-controlled skill sets from ClawHub.

Part 11 — Migrate OpenClaw → Hermes

Hermes ships a native migration path:

hermes claw migrate

This imports OpenClaw workspace layout, channel configuration, and compatible skills where possible.

After migration:

hermes doctor
hermes claw migrate --help # inspect flags
# Compare cron + channel config manually
hermes gateway

Community fallback for older Hermes versions: openclaw-to-hermes.

Side-by-side cutover (recommended for production personal assistants):

Migrate with hermes claw migrate
Run Hermes gateway on a new Telegram bot
Keep OpenClaw on the old bot until Hermes passes your test checklist
Switch DNS/webhooks if applicable
Decommission OpenClaw daemon when satisfied

Part 12 — Security comparison

| Risk | OpenClaw mitigation | Hermes mitigation |
|------|---------------------|-------------------|
| Malicious skill | `openclaw skills verify`, review scripts | Review `SKILL.md` + scripts before enabling |
| Shell/RCE | Docker sandbox (docs strongly recommend) | Remote sandboxes, minimal VPS install (`--skip-browser`) |
| Open gateway | Local bind, auth tokens | `hermes doctor`, pairing/allowlists |
| Prompt injection via chat | Model choice, tool allowlists | Same — use strongest model available |

Rule for both: skills are code. Treat ClawHub and awesome-hermes-agent entries as untrusted until reviewed.

Part 13 — Run both side by side (this repo)

From the repo root:

cd guides/hermes-vs-openclaw
chmod +x verify-comparison.sh
./verify-comparison.sh

Optional full stack:

| Terminal | Command |
|----------|---------|
| A | Start RAG API per [qwen-agentic-rag](https://ayush7614.github.io/agentic-ai-ecosystem/guides/qwen-agentic-rag/) |
| B | `openclaw gateway` (messaging assistant) |
| C | `hermes --tui` (compare tool/skill behavior) |

OpenClaw consumes RAG via the agentic-rag skill. Hermes can call the same HTTP API via a custom skill or MCP wrapper.

Part 14 — Decision guide

| Profile | Pick |
|---------|------|
| Indie hacker, Telegram/WhatsApp only, loves npm | **OpenClaw** |
| ML researcher, multi-agent, Nous ecosystem | **Hermes** |
| Existing OpenClaw user, curious about learning loop | **Hermes** via `hermes claw migrate` |
| Need reproducible skill catalog, not auto-writes | **OpenClaw** + ClawHub |
| Building on this repo's RAG guides | **OpenClaw** primary; Hermes optional second runtime |

You can also run OpenClaw for channels and Hermes for batch/cron evolution against the same RAG API — they are not mutually exclusive at the API layer.

Part 15 — Troubleshooting

| Symptom | OpenClaw fix | Hermes fix |
|---------|--------------|------------|
| CLI not found | `npm i -g openclaw@latest`; check `node -v` | `source ~/.zshrc`; re-run installer |
| Doctor fails | Re-run `openclaw onboard` | `hermes setup --portal` |
| Gateway won't start | `openclaw gateway restart`; check port 18789 | `hermes doctor`; check channel tokens |
| Skills missing | `openclaw skills list`; workspace path | `ls ~/.hermes/skills`; new session |
| Node too old | nvm install 22; [`use-node22.sh`](https://github.com/Ayush7614/agentic-ai-ecosystem/blob/main/guides/openclaw-gemma-rag/use-node22.sh) | N/A |
| Migration incomplete | — | `hermes claw migrate`; compare cron/channels; try [openclaw-to-hermes](https://github.com/0xNyk/openclaw-to-hermes) |
| Both fight for Telegram | Use two bot tokens | Use two bot tokens |

Summary

| Dimension | Winner (typical) |
|-----------|------------------|
| Channel polish + dashboard | OpenClaw |
| Self-improving skills | Hermes |
| npm / TypeScript ecosystem | OpenClaw |
| Multi-agent + research tooling | Hermes |
| Local Gemma + RAG (this repo) | OpenClaw |
| OpenClaw → Hermes migration | Hermes (`hermes claw migrate`) |

Next steps: