The short answer: yes, they port
A skill you install on Hermes will install — byte-identical — on OpenClaw, and vice versa. Both harnesses speak the standard agent-skill format (SKILL.md with YAML frontmatter + supporting examples/, references/, scripts/, templates/ dirs), and both pull from the same registries.
Verified 2026-07-31: the skill agent-self-evaluation (among many others) is installed identically on both — same SKILL.md (7625 bytes, identical content), same directory layout — on this machine at ~/.hermes/skills/ and ~/.openclaw/skills/.
How each harness manages skills
| Hermes | OpenClaw | |
|---|---|---|
| Command | hermes skills | openclaw skills |
| Subcommands | browse, search, install, inspect, list, check, update, audit, uninstall, publish, snapshot, tap, config | list, inspect, install (+ onboard) |
| Bundles | hermes bundles (aliases for multiple skills) | (per-agent workspaces) |
| Self-improvement | hermes curator — background skill maintenance, refines skills from experience | (no equivalent curator) |
| Registries | skills.sh, well-known agent skill endpoints, GitHub, ClawHub, others | shared skill ecosystem |
The format is shared; the runtime isn’t
The skill package is portable (identical format + content). What differs is the runtime that executes it:
- Hermes runs skills inside its Python agent, and a curator background process refines skills from experience (the self-improving loop). A skill can get better as Hermes uses it.
- OpenClaw runs skills via its Node gateway, scoped per-agent workspace. No equivalent self-improving curator.
A skill’s scripts/ may be language-specific (bash/python/node). The SKILL.md instructions and the skill’s contract are harness-agnostic — but if a skill shells out to Python, it expects a Python runtime (which Hermes always has; OpenClaw needs Python available). That’s the main compatibility gotcha, and it’s per-skill, not per-harness.
What this means for AgentHub Phase 2
The original thesis was “build an adapter SDK so one skill runs on both.” The live evidence revises that: portability is already solved by the shared format + registries. The real gap is discovery + quality + cross-harness compatibility notes — skills.sh and ClawHub exist, but finding good skills and knowing which ones have runtime quirks across harnesses is unsolved. AgentHub Phase 2 will focus there: a curated, reviewed directory on top of the shared skill ecosystem, not a portability adapter.
Sources + freshness
The shared-format claim was verified 2026-07-31 by diffing an installed skill (agent-self-evaluation) across ~/.hermes/skills/ and ~/.openclaw/skills/ (byte-identical SKILL.md + layout). CLI subcommands from hermes skills --help and openclaw skills --help. Both projects move fast — re-check before relying on subcommand lists.