Built-in Tools

Every tool OpenClaw ships with — from file operations to browser automation.

Core tools (read, exec, edit, write) are always available, subject to tool policy. TOOLS.md in your workspace doesn't control which tools exist — it's guidance for how you want them used.

Core Tools (Always Available)

apply_patch Tool

Optional tool gated by the tools.exec.applyPatch config flag. Applies a structured patch format for multi-file or multi-hunk edits — more efficient than multiple single edit calls when you need coordinated changes across several files.

openclaw config set tools.exec.applyPatch true

Once enabled, the agent can submit unified-diff-style patches that modify multiple files in a single tool call, reducing round trips and improving atomicity of related changes.

Browser Tools

Lobster

OpenClaw's dedicated browser automation engine. Named after the project mascot 🦞. Lobster handles complex multi-step web workflows — navigating pages, interacting with dynamic elements, filling forms, and extracting structured data from web applications that require JavaScript rendering.

Lobster is purpose-built for scenarios where simple HTTP fetch isn't enough: SPAs, pages behind authentication, and multi-step wizards that require maintaining state across interactions.

Web Tools

Built-in utilities for web interactions without launching a full browser. Fetch URLs, parse HTML, and search the web — lightweight alternatives for pages that don't require JavaScript rendering or complex interaction patterns.

Use Web Tools when you need to grab content from a URL, scrape a static page, or perform a quick web search. For anything that requires clicking, scrolling, or authenticated sessions, reach for Browser or Lobster instead.

LLM Task

Delegate sub-tasks to another LLM call within the current agent run. The agent spawns a focused LLM invocation for a specific job, then incorporates the result back into its workflow.

Useful for:

• Summarizing long content that would consume too much context
• Translating text between languages
• Generating structured data (JSON, YAML) from unstructured input
• Classification and extraction tasks embedded in larger workflows

Elevated Mode

Run commands with elevated (sudo/admin) privileges when needed. Controlled via /elevated directives and subject to allowlist restrictions. Use for operations like installing system packages or modifying protected files.

/elevated on    # enable for next commands
/elevated off   # disable

Elevated mode is opt-in per command group. The agent will not escalate privileges without an explicit directive, and the allowlist ensures only approved commands can run with elevated access.

Thinking Levels

Control the agent's reasoning depth for different tasks:

• Default — standard reasoning, suitable for most tasks
• Extended — deeper analysis for complex tasks like architecture decisions, multi-step debugging, or nuanced code review

Configured via agents.defaults.thinkingLevel in your workspace config, or overridden per-message when you need the agent to think harder about a specific problem.

Reactions

The agent can react to messages with emoji reactions on chat platforms (WhatsApp, Telegram, Discord) to acknowledge receipt before a full response is ready. A lightweight feedback mechanism that lets users know the agent has seen their message and is working on it.

Process Management

Background exec tasks are tracked in memory. Long-running commands auto-background after the yieldMs timeout elapses, allowing the agent to continue working while the process runs. Use the process tool to poll status or clear completed sessions.

openclaw process list              # see all active background sessions
openclaw process clear <sessionId>  # terminate and clean up a session

Tool Policy

Tools respect the agent's tool policy. Configure which tools are allowed, denied, or require confirmation in openclaw.json:

{
  "agents": {
    "defaults": {
      "toolPolicy": {
        "exec": "ask",
        "write": "ask",
        "read": "allow"
      }
    }
  }
}

Policy values: "allow" (no confirmation needed), "ask" (agent requests permission before use), and "deny" (tool is blocked entirely). This gives you fine-grained control over what the agent can do autonomously versus what requires your approval.