OpenClaw Skills: How to Find, Install, and Build Custom Agent Capabilities

Out of the box, an OpenClaw agent can handle conversations, search the web, read and write files, execute shell commands, and connect to messaging platforms. That covers a lot. But the real power comes from skills — modular capability packages that teach your agent how to do specific things.

OpenClaw skills and the ClawHub marketplace are the difference between a general-purpose agent and one that knows your exact tools, workflows, and integrations. This guide covers what skills are, how to find and install them, how to build your own, and the skill stack that powers MrDelegate's autonomous operations.

What OpenClaw Skills Are

A skill is a directory containing a SKILL.md file and any supporting resources. The SKILL.md file contains instructions that the agent reads when it determines the skill is relevant to a task. Think of it as a manual the agent loads on demand — it does not read every skill for every message, just the ones that apply.

Skills can include:

• SKILL.md — the core instruction file, written in markdown, describing what the skill does and how to use it
• scripts/ — executable scripts the skill uses (Python, bash, Node.js)
• references/ — documentation, API specs, or reference material the agent needs
• templates/ — prompt templates, document templates, or output formats

Skills are not plugins in the traditional software sense. They are not compiled extensions that hook into the OpenClaw runtime. They are instruction packages — structured knowledge and tools that the agent can access when needed. This makes them simple to build, easy to modify, and completely transparent. You can read exactly what a skill instructs the agent to do.

The skill discovery mechanism works through the AGENTS.md file and the available skills list injected into the agent's context. The agent scans skill descriptions, determines which one most closely matches the current task, and loads that skill's SKILL.md before proceeding.

Finding Skills on ClawHub

ClawHub is the OpenClaw skills marketplace — a registry where the community publishes, shares, and discovers skills. It operates similarly to npm for Node.js packages: a central registry with versioned packages you can install with a command.

To search ClawHub from your terminal:

clawhub search [query]

For example, searching for CRM skills returns all published skills related to CRM integrations — HubSpot, Salesforce, Pipedrive, Airtable. Each listing shows the skill name, description, author, version, and install count.

To view details on a specific skill before installing:

clawhub info [skill-name]

This shows the full description, what the skill enables, any requirements (API keys, installed tools), and the SKILL.md preview so you can see exactly what instructions the agent will receive.

You can also browse skills through the ClawHub web interface at clawhub.com, which includes category filtering (productivity, integrations, development, content, finance, etc.) and sorting by install count, recent updates, and relevance.

When evaluating a skill before installing, look at:

• Install count — popular skills tend to be more reliable and better maintained
• Last updated — skills that have not been updated in six months may reference outdated APIs or tools
• Author — skills from verified publishers or the OpenClaw core team are generally more trustworthy
• SKILL.md preview — read the actual instructions before installing. If they are vague, poorly written, or reference tools you do not have, look for an alternative

Installing a Skill

Installing a skill from ClawHub is a single command:

clawhub install [skill-name]

This downloads the skill to your OpenClaw skills directory (typically /usr/lib/node_modules/openclaw/skills/ or a configured custom path) and makes it available to your agent.

Some skills require configuration after installation — API keys, endpoint URLs, or preference settings. These are typically documented in the SKILL.md file itself. A well-written skill will tell you exactly what environment variables to set or what configuration files to create.

To update all installed skills to their latest versions:

clawhub update

To update a specific skill:

clawhub update [skill-name]

To see all installed skills and their versions:

clawhub list

After installing a skill, your agent can use it immediately — no restart required. The skill appears in the available skills list on the agent's next turn, and the agent will load it when the task matches the skill's description.

If a skill requires a tool to be installed on the host system (a CLI tool, a Python package, an API key in a specific environment variable), the agent will typically tell you when it first tries to use the skill and the dependency is missing. Fix the dependency, and the skill works.

Building a Custom Skill

The most powerful use of OpenClaw skills is building custom capabilities for your specific tools and workflows. If you use a CRM that does not have a published ClawHub skill, or you have internal systems with custom APIs, or you want your agent to follow a very specific process for a task — you build the skill yourself.

A minimal skill requires only a directory and a SKILL.md file. Here is the structure:

my-skill/
  SKILL.md
  scripts/   (optional)
  references/ (optional)

The SKILL.md file follows a standard format:

Description — a one-line description that the agent uses to determine when to load the skill. Write this carefully — it is the trigger for skill selection. Be specific: "Manage HubSpot CRM contacts and deals via the HubSpot API" is better than "CRM integration."

Prerequisites — list any required tools, API keys, or configuration the skill needs to work.

Instructions — the actual step-by-step guidance for the agent. This is where you write the how. Be explicit about what to do, what endpoints to call, what format to use for output, and what to do when things go wrong.

Examples — example tasks and expected agent behavior. These help the agent understand the skill's scope.

The quality of a custom skill comes down to the quality of its instructions. Well-written skills read like clear documentation written for a capable but uninformed operator — explain the goal, the tools available, the process to follow, and the edge cases to handle.

For skills that use scripts, keep the scripts simple and well-commented. The agent will sometimes read script contents to understand what a tool does. Clear, readable code helps the agent use the skill more effectively.

Once built, test your skill by asking your agent to perform the task the skill is designed for. Check that the agent correctly loads the skill (it will mention it in its reasoning) and follows the instructions as intended. Iterate on the SKILL.md until the agent consistently does what you want.

To publish a skill to ClawHub so others can use it:

clawhub publish

Run this from inside the skill directory. You will need a ClawHub account (free) and a package.json with the skill's name, version, and description.

MrDelegate's Pre-Configured Skill Stack

MrDelegate runs on OpenClaw with a curated skill stack built for autonomous business operations. The skills we use daily:

Core infrastructure skills: SSH management, server health monitoring, git operations, process management. These give the CEO agent and infrastructure agents direct access to production systems with appropriate safety checks built into the skill instructions.

Content skills: Blog publishing workflows, SEO optimization checks, HTML template generation, image handling. Content agents use these to write and deploy articles without manual intervention.

Business operations skills: Email management, calendar integration, customer support triage, invoice processing. These connect the AI agents to the tools that run day-to-day operations.

Research and intelligence skills: Web search orchestration, competitor monitoring, market research compilation. The CEO agent uses these to stay current on market conditions and competitive dynamics.

Developer skills: Codex agent spawning, code review, deployment verification, database query execution. These let the development workflow run autonomously from task definition through deployment verification.

Each of these is either a published ClawHub skill we have configured for our environment, or a custom skill built for our specific tools and processes. The investment in building the right skill stack upfront is what makes autonomous operation possible — the agents know exactly how to use our systems because the skills tell them.

If you are building on OpenClaw, start by identifying the three to five tools your agent needs to interact with most. Build or install skills for those first. Once those work reliably, add the next layer. A well-configured skill stack is the difference between an agent that handles tasks competently and one that actually runs your operations.