01
The wrong question (and the one that actually matters)
Most agent comparisons argue over benchmarks and demo speed. That's the wrong test for a business. The real test is what happens on day 90, unattended: does the agent still know your customers, has it gotten better at the jobs it does daily, does it keep running when a vendor changes a pricing page, and can you move it if you have to. OpenClaw, Hermes Agent, and Claude Code are all strong — but they make very different trade-offs on exactly those questions. Pick wrong and the cost isn't a worse demo; it's a rebuild six months in, or a bill that jumps overnight. Five things decide it:
- Self-generating skills — does it learn new abilities from work it already did, or do you hand-write every one?
- Memory + identity — does it remember across sessions, and is its personality/voice pinned in a file you control?
- Where it runs — your own infrastructure, or someone's subscription and hosted backend?
- Openness — open-source licence you can fork and self-host, or proprietary?
- Lock-in risk — if the vendor changes the rules, what breaks, and how fast can you switch?
02
What each one actually is (verified, not vibes)
Quick, accurate framing before the table — each claim here is from the project's own repo or docs, attributed to the right project.
- OpenClaw — open-source (MIT), Node/TypeScript, "a personal AI assistant you run on your own devices" (its repo). Gateway-first: one long-running Gateway process is the single source of truth for sessions and channels, so it meets customers on WhatsApp, Telegram, Slack, Discord, iMessage and more. Skills live as SKILL.md files you author; the community registry is ClawHub.
- Hermes Agent — open-source (MIT), by Nous Research, "the agent that grows with you." Its docs describe autonomous skill creation and skill self-improvement during use, a global SOUL.md for identity, a layered persistent memory, and running on your own infra across local/Docker/SSH/Modal backends, with any OpenAI-compatible provider.
- Claude Code — Anthropic's agentic coding tool. Proprietary. Its docs say most surfaces need a Claude subscription or Anthropic Console account (Terminal and VS Code can also use third-party providers). Memory is CLAUDE.md plus auto memory it writes itself. Polished operator/coder copilot, deepest model — but hosted and subscription/console-billed.
03
The decision table
Same five questions, side by side. Every cell is attributed to the right project from its primary source. No single tool wins every row — that's the point.
| What matters | OpenClaw | Hermes Agent | Claude Code |
|---|---|---|---|
| Self-generating skills | No — skills are SKILL.md files you author manually | Yes — autonomous skill creation + self-improvement during use (its docs) | No auto-authoring — you create skills/commands; it does write its own auto memory |
| Memory + identity | Workspace memory; a SOUL.md persona (workspace-scoped) | Layered persistent memory + a global SOUL.md identity file | CLAUDE.md (you write) + auto memory (it writes), per project |
| Where it runs | Your own devices; one Gateway process | Your own infra — local/Docker/SSH/Modal | Anthropic-hosted models; routines run on Anthropic-managed infra |
| Openness / licence | Open-source, MIT | Open-source, MIT | Proprietary (Anthropic) |
| Model providers | Bring your own model/keys | Any OpenAI-compatible endpoint, bring your own keys | Claude by default; Terminal/VS Code can use 3rd-party providers |
| Best at | Messaging-first assistant across every channel customers use | Running unattended, learning on the job, keeping memory + identity | Hands-on coding/operator copilot with the deepest model |
04
The trade-offs nobody puts on the landing page
Each strength has a tax. Going in eyes-open is the whole job of this guide.
- OpenClaw's big skill registry (ClawHub) is also its biggest risk surface. In early 2026 a supply-chain campaign ("ClawHavoc") seeded ClawHub with hundreds of malicious skills — security researchers (Koi Security, Trend Micro) found malicious instructions hidden inside SKILL.md files, with payloads including a macOS infostealer. The fix is discipline: pin and review every skill you install, don't auto-trust the registry.
- Hermes self-generates skills and remembers everything — which means it accumulates state you have to govern. Self-improvement is a feature, but you still own backups, the SOUL.md, and review of what it taught itself. It's also more involved to stand up than a one-command toy.
- Claude Code is the most polished and the deepest model — but it's proprietary and tied to a subscription or Anthropic Console. In April 2026 Anthropic blocked flat-rate Claude Pro/Max plans from running third-party agents, moving that usage to pay-as-you-go; heavy users reported bills jumping up to ~50× (reported across multiple outlets). Great tool — just don't wire a business's unattended automation to one subscription policy.
05
Pick in 30 seconds
Match the job to the tool. Most serious setups end up using two of these for different layers, not one for everything.
- You want an agent that RUNS the business unattended, gets better at daily jobs, and keeps its memory + identity on infra you own and any model you can swap → Hermes Agent.
- You want a messaging-first assistant that meets customers on every channel they already use, fast to stand up, with a big skill library — and you'll review what you install → OpenClaw.
- You want the best hands-on coding/operator copilot with the deepest model, and you accept proprietary + subscription/console billing → Claude Code.
- Whichever you pick: own the layer underneath it — your API keys, your infra, and the router/brand/billing on top — so a vendor's pricing page can't hold your business hostage.
06
Your 6-point pre-commit checklist
Run this before you build a business on any agent — open-source or not. Each 'no' is a future migration you're signing up for.
- Can I self-host it on infra I control, or does the heavy lifting run on the vendor's servers?
- Is the licence open (can I fork it) or proprietary (I'm a tenant)?
- Can I point it at my own API keys / swap the model behind it without a rebuild?
- Does its memory + identity live in files I own (and back up), not a vendor database?
- If it has a skill marketplace, do I pin and review every skill — or auto-trust strangers' code?
- If the vendor changed pricing tomorrow, what breaks, and how many days to switch?
Get the next drop
New AI build guides + the occasional bonus template. No spam, unsubscribe anytime.
You're in — check your inbox to confirm.
Frequently asked questions
OpenClaw vs Hermes — which is better for running an AI business?
Different shapes. <strong>OpenClaw</strong> is gateway-first: one Gateway process meeting customers across WhatsApp, Telegram, Slack, Discord, iMessage and more, with a large (manually-authored) skill library on ClawHub. <strong>Hermes Agent</strong> is built to run unattended and improve itself — its docs describe autonomous skill creation, a global SOUL.md identity, layered persistent memory, and running on your own infra (local/Docker/SSH/Modal). If you want broad messaging reach now, OpenClaw; if you want an agent that learns on the job and keeps its memory + identity, Hermes.
Is Hermes Agent really self-improving, or is that marketing?
Per Nous Research's own docs, Hermes does autonomous skill creation and "skill self-improvement during use" — it turns successful workflows into reusable skills and curates its own memory across sessions. That's a real architectural feature, not just a tagline. You still own governance: back up the skills it writes and review what it taught itself.
What is SOUL.md, and which agents have it?
SOUL.md is a plain-text file that pins an agent's identity, voice and defaults so they persist across sessions. <strong>Hermes</strong> has a <em>global</em> SOUL.md (its docs call it the agent's default voice). <strong>OpenClaw</strong> also ships a SOUL.md persona, scoped to the workspace. <strong>Claude Code</strong> doesn't use SOUL.md — its equivalent is CLAUDE.md (instructions you write) plus auto memory it builds itself.
Are OpenClaw and Hermes open source? Is Claude Code?
OpenClaw and Hermes Agent are both <strong>open-source under the MIT licence</strong> (their repos), so you can self-host and fork them. <strong>Claude Code is proprietary</strong> — Anthropic's docs note most surfaces require a Claude subscription or Anthropic Console account, though the Terminal CLI and VS Code can use third-party providers.
What's the real lock-in risk with each one?
With OpenClaw and Hermes the lock-in is low — open licence, your infra, bring-your-own-keys — but you own security (e.g. vetting ClawHub skills after the 2026 supply-chain attack) and operations. Claude Code's risk is pricing/policy: in April 2026 Anthropic blocked flat-rate Claude Pro/Max plans from running third-party agents and moved that usage to pay-as-you-go, with some heavy users reporting up to ~50× cost increases. The hedge for all three: own your keys, your infra, and the orchestration layer on top.
Sources · OpenClaw — official GitHub repository (MIT, SKILL.md skills, Gateway, SOUL.md) · Hermes Agent Documentation — Nous Research (self-improving skills, global SOUL.md, own-infra, providers) · Claude Code Overview — Anthropic docs (subscription/Console, CLAUDE.md + auto memory, third-party providers) · Anthropic blocks OpenClaw from Claude subscriptions in cost crackdown — TNW · Anthropic temporarily banned OpenClaw's creator from accessing Claude — TechCrunch · Malicious OpenClaw skills used to distribute Atomic macOS Stealer — Trend Micro · Hundreds of malicious skills found in OpenClaw's ClawHub — eSecurity Planet