OpenClaw Tool Calling Explained — How Skills and Tools Work

What Is Tool Calling in OpenClaw?

Tool calling is the mechanism that lets OpenClaw's agent actually DO things rather than just talk about them. When your agent needs to search the web, send a Slack message, execute a shell command, or call an API, it uses tools. Understanding how this works makes you dramatically better at configuring and extending OpenClaw.

Built-In Tools vs Skill Tools

OpenClaw has two categories of tools:

Built-in tools are core capabilities always available to your agent:

• exec — run shell commands (with sandbox controls)
• read/write/edit — file system operations in your workspace
• browser — control a real browser for web automation
• message — send/receive messages on connected channels
• web_search — search the web via Brave API
• web_fetch — fetch and parse web page content

Skill tools are defined in SKILL.md files. Each skill describes the tools it provides and how to use them:

# Example SKILL.md — GitHub Skill
## Tool: create_issue
POST https://api.github.com/repos/{owner}/{repo}/issues
Headers: Authorization: Bearer $GITHUB_TOKEN
Body: {"title": "$TITLE", "body": "$BODY", "labels": ["$LABEL"]}

## Tool: list_prs
GET https://api.github.com/repos/{owner}/{repo}/pulls?state=open

How the Agent Decides to Use a Tool

When you give your agent a task, it reasons about which tools to use — in what order and with what parameters. This happens through the LLM's function-calling capability. Your AGENTS.md can guide this reasoning:

## Tool Use Guidelines
- Always check workspace files before searching the web
- For external API calls: verify credentials exist before calling
- On exec tool: prefer read-only commands unless write is explicitly needed
- After any external action: log the result to workspace if it might be needed later

Tool Call Approval Model

By default, certain tools require approval before execution. You control this in your configuration:

# In openclaw.json:
{
  "permissions": {
    "exec": "ask",           // Always ask before running shell commands
    "web_search": "allow",   // Always allow web search
    "message": "allow",      // Allow sending messages
    "browser": "ask"         // Ask before browser automation
  }
}

Building Custom Tool Definitions

Any REST API becomes a tool through SKILL.md. The pattern is:

1. Define the tool name and description
2. Specify the HTTP method, URL, headers, and body format
3. Reference env variables for credentials
4. Add example usage so the agent knows when to invoke it

Tool Chaining

Your agent can chain tools across a workflow. Example: web search → extract data → write to file → send Slack message. Each step uses a different tool, and the agent coordinates the output of one as input to the next.

## Workflow Example (AGENTS.md)
For competitor research:
1. web_search: "[competitor] new features 2026"
2. web_fetch: top 3 result URLs
3. write: save summary to research/competitors/[name]-[date].md
4. message: post summary to #competitive-intel

Ready to unlock this for your workflow? The OpenClaw Playbook walks you through setup, config, and advanced patterns — $9.99, one-time.