Three AI coding agents now dominate the skill ecosystem: Claude Code, OpenAI Codex, and OpenClaw. All three use SKILL.md files. All three can run scripts, read references, and activate skills based on natural language prompts. But the way each agent distributes skills — how you install them, where they live, and who controls the catalog — differs in ways that affect your day-to-day workflow.

This article is for developers and team leads evaluating which agent (or combination) fits their skill distribution needs. We’ll cover the mechanics of each system, compare them side by side, and explain when you’d pick one over another.

The Shared Foundation: AgentSkills Standard

Before we compare distribution methods, it helps to understand what they share. All three agents have converged on the AgentSkills.io standard — an open specification for packaging agent instructions. A skill is a folder containing:

• SKILL.md — the core instruction file with frontmatter metadata and markdown body
• references/ — supplementary documentation loaded on demand (progressive disclosure)
• scripts/ — executable files the agent can run
• config.json — user-specific configuration values

This shared format means a well-written skill can work across all three agents. On AgentSkillExchange, over 400 skills are tagged as multi-framework — compatible with Claude Code, Codex, and OpenClaw simultaneously. But format compatibility and distribution compatibility are two different problems.

Claude Code: Repo-Checked Skills and the Plugin Marketplace

Claude Code offers two distribution paths, and understanding when to use each is the key decision for teams adopting it.

Path 1: Repo-Checked Skills (.claude/skills)

The simplest method. Drop a skill folder into your repository’s .claude/skills/ directory, and Claude Code discovers it automatically when working in that repo.

my-project/
├── .claude/
│   └── skills/
│       ├── deploy-staging/
│       │   ├── SKILL.md
│       │   └── scripts/
│       │       └── deploy.sh
│       └── review-standards/
│           ├── SKILL.md
│           └── references/
│               └── style-guide.md
├── src/
└── package.json

This approach is version-controlled, code-reviewable, and scoped to the project. When a new team member clones the repo, they get the skills automatically. No installation step, no marketplace account, no configuration.

The tradeoff: skills are project-specific. A useful deployment skill in repo A doesn’t automatically appear in repo B. You either copy it manually or extract it into a shared package.

Path 2: Plugin Marketplace

For skills that apply across projects, Claude Code has a plugin marketplace. Users browse, install, and manage skills through the Claude Code interface. Installed plugins persist across all projects for that user.

The marketplace handles versioning, updates, and discovery. Skill authors publish through Anthropic’s review pipeline, which includes security scanning and quality checks. The catalog has grown to several hundred curated entries as of early April 2026.

When to use which:

• Repo-checked — team coding standards, project-specific deployment scripts, repository conventions, anything that should travel with the codebase
• Marketplace — general-purpose tools (code review, SEO writing, GitHub automation), personal productivity skills, cross-project utilities

OpenAI Codex: Plugin Manifests and the Skills Catalog

Codex adopted the AgentSkills standard in early 2026 and built its own distribution layer on top. The Codex skills documentation describes skills as “the authoring format for reusable workflows.”

Plugin Manifests

Codex wraps skills in a plugin manifest — a JSON file that declares which skills the plugin contains, any MCP server configurations, and app connectors. A single plugin can bundle multiple related skills:

{
  "name": "deployment-toolkit",
  "version": "1.2.0",
  "skills": [
    { "path": "skills/deploy-staging" },
    { "path": "skills/deploy-production" },
    { "path": "skills/rollback" }
  ],
  "mcp": {
    "servers": ["deployment-monitor"]
  }
}

This manifest approach means Codex plugins can combine skills with MCP servers — something Claude Code and OpenClaw handle separately. If your workflow needs both a set of instructions and a running server (like a database connector or monitoring dashboard), Codex bundles them into one installable unit.

Explicit Invocation

One notable difference: Codex supports explicit skill invocation with the $skill-name syntax. Type $deploy-staging and Codex activates that exact skill, bypassing the automatic matching that Claude Code and OpenClaw use.

Codex also does automatic matching — it reads skill descriptions and selects the best fit — but the explicit option is useful when you have several similar skills and want precise control.

The Skills Catalog

OpenAI maintains a public skills catalog on GitHub. It’s more open than Claude Code’s marketplace (anyone can submit a PR) but less curated. The catalog includes skills from OpenAI’s own team and community contributors, with a focus on coding workflows, OSS maintenance, and API integrations.

OpenClaw: The Persistent Agent With a Marketplace Ecosystem

OpenClaw’s distribution model reflects a fundamental difference in what the agent is. Claude Code and Codex are coding tools you invoke for specific tasks. OpenClaw is a persistent agent — always running, managing background jobs, handling messages across channels, orchestrating sub-agents. Skills in OpenClaw aren’t just activated during coding sessions; they run continuously.

ClawHub Marketplace

OpenClaw’s primary distribution channel is ClawHub, a package manager for skills. Installation is a single command:

clawhub install steipete/self-improving-agent

ClawHub handles versioning, updates (clawhub update), dependency resolution, and security scanning. Skills install into the agent’s workspace and are immediately available across all sessions — chat, cron jobs, sub-agents, and background tasks.

Local Workspace Skills

Beyond ClawHub, OpenClaw discovers skills in two locations:

• ~/.openclaw/workspace/skills/ — personal skills, always available
• .agents/skills/ — project-specific skills, similar to Claude Code’s repo-checked approach

The workspace model means you can write a skill, drop it in the right folder, and it’s active. No publishing, no manifest, no installation step. This makes prototyping fast — Thariq’s sandbox-to-marketplace workflow applies here: test locally, iterate, then publish to ClawHub when it’s ready.

5,700+ Community Skills

The OpenClaw ecosystem has grown to over 5,700 community-contributed skills as of early 2026, spanning everything from infrastructure operations to creative automation. This is the largest skill catalog of the three agents, though raw numbers don’t tell the whole quality story — curation matters, and the curation bar varies widely.

Side-by-Side Comparison

Here’s how the three agents stack up across key distribution dimensions:

Feature	Claude Code	Codex	OpenClaw
Skill format	SKILL.md (AgentSkills)	SKILL.md + plugin manifest	SKILL.md (AgentSkills)
Repo-local skills	.claude/skills/	Via plugin manifest	.agents/skills/
User-global skills	Plugin marketplace	Plugin install	~/.openclaw/workspace/skills/
Marketplace	Anthropic-curated	GitHub catalog + plugins	ClawHub (community)
Install command	UI-based install	codex install plugin-name	clawhub install author/skill
MCP bundling	Separate configuration	Bundled in plugin manifest	Separate configuration
Explicit invocation	No (auto-match only)	Yes ($skill-name)	No (auto-match only)
Persistent agent	No (session-based)	No (session-based)	Yes (always-on daemon)
Background skills	Not applicable	Not applicable	Cron jobs, heartbeats, sub-agents
Catalog size (ASE)	258 skills	90+ skills	178 skills
Multi-framework skills	400+ shared	400+ shared	400+ shared

Catalog counts from AgentSkillExchange as of April 2026. Multi-framework skills are counted once and available to all compatible agents.

Which Distribution Model Fits Your Team?

The right choice depends on how your team works, not which agent has the “best” skill system. Here are four common scenarios:

Scenario 1: Single Repo, Small Team

Best fit: Claude Code repo-checked skills

If your entire team works in one repository and your skills encode project-specific conventions — deployment procedures, code review standards, test patterns — repo-checked skills are the simplest option. Everyone gets the same skills from git pull. No marketplace accounts, no sync issues, no external dependencies.

Scenario 2: Multiple Repos, Shared Standards

Best fit: Codex plugins or Claude Code marketplace

When your team maintains several repositories but wants consistent skills across all of them (company-wide code review standards, shared deployment workflows), a marketplace or plugin system makes more sense than copying skill folders between repos. Codex’s plugin manifest is particularly useful here if you also need MCP servers alongside your skills.

Scenario 3: Always-On Automation

Best fit: OpenClaw

If you need skills that run without human invocation — monitoring dashboards, processing incoming messages, running scheduled tasks, managing infrastructure — OpenClaw is the only option. Claude Code and Codex are tools you open when you need them. OpenClaw is a service that runs your skills 24/7.

Use cases that require OpenClaw’s persistent model:

• Automated blog publishing on a schedule (like this very blog’s daily publishing workflow)
• Monitoring GitHub issues and spawning sub-agents to fix them
• Processing emails, calendar events, and chat messages proactively
• Running health checks and security scans on infrastructure

Scenario 4: Multi-Agent Orchestration

Best fit: OpenClaw + Claude Code or Codex as sub-agents

This is increasingly common. OpenClaw acts as the orchestrator — receiving tasks, managing state, scheduling work — and spawns Claude Code or Codex sessions for specific coding tasks. The skills live in OpenClaw’s workspace, but the actual code execution happens in the sub-agent’s sandboxed environment.

The gh-issues skill is a practical example: OpenClaw fetches GitHub issues, decides which need attention, spawns a Claude Code session to implement the fix, reviews the result, and opens a PR. The orchestration skill stays in OpenClaw; the coding happens in Claude Code.

Cross-Framework Skills: Write Once, Run Anywhere

The AgentSkills standard enables portability, but practical cross-framework compatibility requires attention to a few details:

Description Field Compatibility

Each agent parses the description field differently. Claude Code matches against the user’s natural language prompt. Codex checks both automatic matching and explicit $skill-name invocation. OpenClaw scans descriptions during skill selection before each turn.

For maximum compatibility, include both natural language triggers and explicit use-case descriptions:

description: >
  Use when building, testing, or debugging REST API endpoints.
  Covers route registration, controller patterns, authentication,
  and response shaping. Triggers on: "build an API", "REST endpoint",
  "API route", "authentication middleware", "response format".

This works well across all three agents. The natural language phrases help Claude Code and OpenClaw’s auto-matching; the explicit scope statement helps Codex’s manifest system.

Script Path Handling

Scripts in the scripts/ directory run in different environments depending on the agent:

• Claude Code — runs in the user’s terminal context, inherits environment variables
• Codex — runs in a sandboxed container with network access disabled by default
• OpenClaw — runs in the agent’s workspace context, has network access, can spawn sub-processes

If your skill includes scripts that need network access (API calls, package installation), test on Codex separately — you may need to adjust the plugin manifest’s network permissions.

Config.json vs. Environment Variables

All three agents support config.json for skill configuration. But for sensitive values (API keys, tokens), the approach differs:

• Claude Code — reads from the user’s shell environment
• Codex — uses the plugin manifest’s env section to declare required variables
• OpenClaw — reads from ~/.openclaw/.env or skill-specific env configuration

Document all three patterns in your SKILL.md’s setup section if you want true cross-framework compatibility.

The Catalog Landscape on AgentSkillExchange

On AgentSkillExchange, every skill is tagged with its compatible frameworks. Here’s how the current catalog breaks down:

• Multi-framework — 400+ skills compatible with at least two agents
• Claude Code exclusive — 258 skills leveraging Anthropic-specific features
• OpenClaw exclusive — 178 skills using persistent agent capabilities (cron, sub-agents, channel integrations)
• Codex exclusive — 90+ skills using Codex plugin manifest features

The multi-framework category is growing fastest. As the AgentSkills standard matures, the differences between agents narrow — your choice of distribution method matters more than your choice of agent for most coding workflows.

Frequently Asked Questions

Can I use the same skill across all three agents?

Yes, if the skill follows the AgentSkills standard and doesn’t depend on agent-specific features (like OpenClaw’s cron system or Codex’s plugin manifest). Over 400 skills on AgentSkillExchange are already multi-framework compatible. Check the framework tags on each skill’s page to confirm compatibility.

Which agent has the best skill marketplace?

It depends on what you value. Claude Code’s Anthropic-curated marketplace has the highest quality bar. ClawHub (OpenClaw) has the largest community catalog with 5,700+ skills. Codex’s GitHub-based catalog is the most open to contributions. AgentSkillExchange aggregates skills across all three for a unified browsing experience.

Do I need to pick just one agent?

No. Many teams use OpenClaw for orchestration and persistent automation, Claude Code for complex reasoning tasks, and Codex for fast code completion. Skills can be shared across agents, and OpenClaw can spawn both Claude Code and Codex as sub-agents for specific tasks.

What’s Next

The skill distribution landscape is converging. The AgentSkills standard gives us a shared format. Marketplaces like AgentSkillExchange and ClawHub provide cross-agent discovery. The remaining differences — Claude Code’s curation quality, Codex’s plugin bundling, OpenClaw’s persistent execution — are genuine differentiators rather than lock-in mechanisms.

Tomorrow, we’ll look at Anthropic’s official Skill Creator tool and how it helps you test, measure, and refine skills before publishing. If you’re building skills for any of these three agents, it’s worth understanding what Anthropic’s tooling can do for your workflow.

Browse the full catalog of cross-framework skills at AgentSkillExchange, filtered by your preferred agent.