Will a Claude Code skill work on Hermes Agent or OpenClaw?
Short answer: mostly yes. Claude Code, Hermes Agent and OpenClaw all read the same SKILL.md format, and all 142 vetted skills we checked (from 23 repos) had the two required fields, a lowercase hyphenated name that matched the folder, and a description in the length limit except one (141 of 142). 65 of 142 (45.8%) need nothing else. The rest need a look: 23 use Claude Code-only frontmatter or tools, 40 need an API key that none of them declares in a way Hermes or OpenClaw reads, and 17 ship scripts that need a terminal.
We run Agent Coach, which recommends skills and MCP servers to agents and security-scans each one at a pinned commit. Agent Coach supports Hermes Agent and OpenClaw as agent platforms, while many skills are written and tested on Claude Code. So on 2026-11-20 we pulled every vetted single-folder skill in our library and read its SKILL.md at the exact commit we scanned. That's 142 of the 198 vetted skill entries; the others are whole packs, our own text-only items, or skills whose folder we couldn't match at the scanned commit. Format rules below come from the Hermes skills docs, the OpenClaw skills docs and the agentskills.io standard.
The shared core: 141 of 142 pass every check
| Check (agentskills.io core) | Skills passing | Share |
|---|---|---|
Has name and description in YAML frontmatter | 142 of 142 | 100.0% |
name is lowercase-hyphenated, 64 characters or fewer | 142 of 142 | 100.0% |
name matches the folder name | 142 of 142 | 100.0% |
description is 1,024 characters or fewer | 141 of 142 | 99.3% |
This is the part all three agents agree on. Hermes says its skills "are compatible with the agentskills.io open standard"; OpenClaw says it "follows the AgentSkills spec" and needs only name and description. OpenClaw's node-hosted mode also requires the folder name to match name, and every skill we checked already did. The one long description (1,068 characters) would most likely still load; agents trim descriptions to save context, OpenClaw included.
What doesn't carry over
| Pattern | Skills | Share | What happens off Claude Code |
|---|---|---|---|
allowed-tools in frontmatter | 12 | 8.5% | Neither the Hermes nor the OpenClaw docs list it. Don't count on the tool limit it sets |
argument-hint | 11 | 7.7% | Claude Code slash-command hint; not documented elsewhere |
context / agent (run in a subagent) | 2 | 1.4% | Claude Code subagent routing; the skill runs inline instead |
disable-model-invocation | 12 | 8.5% | OpenClaw honors it. The Hermes skills page doesn't document it, so test whether Hermes loads the skill on its own |
user-invocable | 9 | 6.3% | OpenClaw honors it. Hermes turns every installed skill into a slash command anyway |
| Body tells the agent to call the Skill tool | 9 | 6.3% | Hermes and OpenClaw load skills their own way; the agent has to work out what "call the Skill tool" means |
Body mentions CLAUDE.md or a .claude/ path | 37 | 26.1% | Mostly as one of several places to look for a context file. Harmless, but the file won't be there unless you add it |
Counting only the patterns that change behavior (the Claude-only keys plus calls to Claude's tools), 23 of 142 skills (16.2%) are Claude-specific. They cluster in a few repos: one popular engineering pack chains its skills through the Skill tool, and a web-research vendor sets allowed-tools, argument-hint and user-invocable on 8 of its skills. The fix is small: rewrite "call the Skill tool with X" as "load the X skill", and accept that the tool allowlist won't apply.
The bigger gap: undeclared keys
40 of 142 skills (28.2%) mention an API key, token or secret by environment-variable name. None of them (0) declare it in a field Hermes or OpenClaw reads. That matters because both agents have a place for it:
- Hermes:
required_environment_variablesmakes Hermes ask for the key securely when the skill loads in the local CLI, and passes it toterminalandexecute_codesandboxes. Messaging surfaces never ask for secrets in chat. - OpenClaw:
metadata.openclaw.requires.env(andrequires.bins) letopenclaw skills checktell you a skill isn't ready before your agent tries it. A skill with nometadata.openclawblock is always eligible, so OpenClaw shows it as usable even when the key is missing.
Only 2 of 142 skills have a metadata.openclaw block at all (one declares a required binary, the other only an emoji and homepage). 0 have a metadata.hermes block and 0 use the Hermes platforms OS filter. In practice that means: before you install a skill on Hermes or OpenClaw, search its SKILL.md for _API_KEY, _TOKEN and _SECRET, put the key where your agent keeps secrets, and add the declaration yourself if you maintain a fork. Our guide on storing API keys for skills and MCP servers covers where.
Scripts need a terminal, and a read
17 of 142 skills (12.0%) ship a scripts/ folder; almost half are document and testing skills from one official repo. On Hermes, a skill that runs scripts needs the terminal toolset, and authors can say so with metadata.hermes.requires_toolsets: [terminal] so the skill hides itself when the tool isn't there. None of these 17 declare it. On any agent, scripts run with your agent's permissions, so read them before the first run. Hermes scans hub installs for exfiltration, prompt injection and destructive commands, and --force can't override a dangerous verdict, but a scanner is a floor, not a review. See how to vet an MCP server or agent skill.
A 4-step check before you move a skill
- Read the frontmatter. If it only has
name,descriptionand maybelicenseormetadata, it will load anywhere. Note anyallowed-tools,contextordisable-model-invocation. - Search the body for Claude's tools. "Skill tool", "AskUserQuestion", "TodoWrite", "$ARGUMENTS". Rewrite them as plain instructions.
- List the keys and binaries. Search for
_API_KEY/_TOKENand for CLIs it tells the agent to run. Set them up, and declare them in your agent's format. - Install into a throwaway profile first. Hermes:
hermes profile create test --no-skills, then install the one skill. OpenClaw: a separate--profile, thenopenclaw skills check.
Where to get skills for Hermes and OpenClaw
Hermes can install from skills.sh, ClawHub, LobeHub, GitHub repos and any site that publishes /.well-known/skills/index.json. OpenClaw installs from ClawHub. Our library lists each skill's install command and scan result at a pinned commit; 198 skill entries currently pass. For how sources compare, see where to find agent skills and MCP servers; for re-checking after install on OpenClaw, see how to review a ClawHub skill.
Agent Coach itself ships as one portable SKILL.md. On Hermes: hermes skills install well-known:https://upskill.coach/.well-known/skills/agent-coach. On OpenClaw: openclaw skills install @f-o-x11/agent-coach. Or paste one line into any agent:
Read https://upskill.coach/start?ref=blog-skill-portability and follow it.It sets one goal for your agent, recommends vetted skills and MCP servers for it, and gives each installed item a keep, temporary or remove verdict from daily check-ins. Developer docs: https://upskill.coach/docs.
What we have not measured
We read files; we didn't install all 142 skills on Hermes and OpenClaw and run them. "Not documented" means the key isn't on the Hermes or OpenClaw skills page as of 2026-11-20; an agent may still tolerate or ignore it. Our pattern matching (Claude tool names, environment-variable names) can miss renamed tools or keys described in prose. The set is our vetted library, which leans toward developer and go-to-market skills, not a random sample of all skills. Per-skill results are in our build data; ask us for the list.
FAQ
Can Hermes Agent use skills written for Claude Code?
Usually, yes. Hermes reads the same SKILL.md format (the agentskills.io standard): a folder with a SKILL.md that has a name and description. All 142 vetted single-folder skills we checked had both, with the name matching the folder. What doesn't carry over is Claude Code-only frontmatter like allowed-tools and instructions that call Claude's own tools, which 23 of them (16.2%) use.
Can OpenClaw use Claude Code skills?
Yes for the same reason: OpenClaw follows the AgentSkills spec and needs only name and description. OpenClaw also documents user-invocable and disable-model-invocation, so those two keys keep working there. A skill without a metadata.openclaw block is always eligible, which means OpenClaw can't warn you about a missing binary or API key unless the author declared it.
How do I install a Claude Code skill on Hermes?
Hermes can install from skills.sh, ClawHub, GitHub repos and well-known endpoints, for example hermes skills install skills-sh/<owner>/<repo>/<skill> or hermes skills install well-known:https://site/.well-known/skills/<name>. Every hub install is security-scanned; community sources can need --force for warn-level findings, and a dangerous verdict can't be forced. You can also copy one skill folder into ~/.hermes/skills/.
What should I change in a skill before using it on another agent?
Three things. Replace calls to Claude-only tools (the Skill tool, AskUserQuestion) with plain instructions. Declare the API keys it needs where your agent reads them (Hermes required_environment_variables, OpenClaw metadata.openclaw.requires.env). And if it ships scripts, make sure your agent has a terminal tool and you have read the scripts.
Does a portable skill mean a safe skill?
No. Portability is about format. Safety is about what the skill tells your agent to do. We security-scan every item in our library at a pinned commit; see the scan results. A skill can load cleanly on three agents and still send data somewhere you didn't expect.
Questions or corrections: ops@avatar33.com.
Published 2026-11-20, updated 2026-11-20 by Avatar 8. Markdown copy: /blog/will-a-claude-code-skill-work-on-hermes-or-openclaw.md