OpenClaw Tools

🔧 Built-In Tool Categories

OpenClaw exposes typed tools directly to the agent — no shelling out required. Tools are organized into logical groups.

📁

File System

Read, write, and edit files in the workspace. Supports text and images.

read write edit apply_patch

group:fs

⚡

Shell & Runtime

Execute commands, manage background processes, and run scripts.

exec process

group:runtime

🌐

Web Research

Search the web and fetch content from URLs for research and analysis.

web_search web_fetch

group:web

🖥️

Browser Automation

Control a real browser for web automation, screenshots, and UI interaction.

browser canvas

group:ui

💬

Sessions & Agents

List sessions, send messages between agents, and spawn sub-agents.

sessions_list sessions_send sessions_spawn session_status

group:sessions

🧠

Memory & Recall

Semantic search across memory files for context and continuity.

memory_search memory_get

group:memory

⏰

Automation & Cron

Schedule tasks, manage cron jobs, and control the gateway.

cron gateway

group:automation

📨

Messaging

Send messages across channels (Telegram, Discord, WhatsApp, etc.).

message

group:messaging

📡

Device Control

Control paired devices — camera, screen capture, notifications, and remote execution.

nodes

group:nodes

🎚️ Tool Profiles

Control which tools your agent can access with built-in profiles.

🔒

minimal

Session status only. For read-only agents.

session_status

💻

coding

File system, runtime, sessions, memory, and images. Perfect for coding agents.

group:fs, group:runtime, group:sessions, group:memory, image

💬

messaging

Message tools plus session management. For chat-focused agents.

group:messaging, sessions_*, session_status

🔓

full

No restrictions. All tools available (default when unset).

* (all tools)

openclaw.json — Set a tool profile

{
  "tools": {
    "profile": "coding",
    "deny": ["exec"]  // Optional: deny specific tools
  }
}

🔍 Key Tools Explained

Deep dive into the most powerful tools in OpenClaw's arsenal.

🖥️

browser

Full browser automation with Playwright-style actions

Core Actions

status Check browser state

start/stop Control browser lifecycle

snapshot Get accessibility tree or AI description

screenshot Capture visible page

act Click, type, press, hover, drag

navigate Go to URL

// Typical browser automation flow
browser → status/start
browser → snapshot (get page structure)
browser → act (click element by ref)
browser → screenshot (verify result)

⚡

exec

Run shell commands with background support and PTY

✓ Background execution — Commands auto-background after yieldMs

✓ PTY support — Run interactive terminal programs

✓ Elevated mode — Run on host with elevated permissions

✓ Node targeting — Execute on remote macOS nodes

// Key parameters
{
  "command": "npm run build",
  "workdir": "/project",
  "yieldMs": 10000,      // Background after 10s
  "timeout": 1800,       // Kill after 30 min
  "pty": true,           // For interactive CLIs
  "host": "node",        // Run on remote node
  "node": "office-mac"   // Target specific node
}

⏰

cron

Schedule tasks, reminders, and recurring jobs

Schedule Types

at One-shot at specific time

every Recurring interval (e.g., every 30 min)

cron Standard cron expression

// Example: Daily 9am reminder
{
  "name": "morning-check",
  "schedule": { 
    "kind": "cron", 
    "expr": "0 9 * * *",
    "tz": "Australia/Brisbane"
  },
  "payload": { 
    "kind": "systemEvent", 
    "text": "Morning check-in time!" 
  },
  "sessionTarget": "main"
}

🤖

sessions_spawn

Spawn sub-agents for parallel work

✓ Isolated sessions — Sub-agent runs in its own context

✓ Non-blocking — Returns immediately, announces when done

✓ Model override — Use different model per task

✓ Cleanup options — Keep or delete session after

// Spawn a research sub-agent
{
  "task": "Research the latest AI agent frameworks and summarize the top 5",
  "label": "research-ai-frameworks",
  "model": "claude-opus-4-6-thinking"
}

📌 What Actually Matters in OpenClaw Tools in 2026

The docs and repo make the direction pretty clear: OpenClaw is optimizing for grounded, recoverable agent work — not just flashy demos.

🌐

Web Research Got Smarter

web_search now supports multiple providers, not just one
Brave, Gemini, Grok, Kimi, and Perplexity are all documented options
web_fetch is still the lightweight HTTP reader — great for docs, bad for JS-heavy apps
For login flows and real UI work, browser is still the right move

🧙

The Wizard Is the Default Path

openclaw onboard is the recommended setup flow
It now handles gateway setup, channels, daemon install, skills, and web search configuration in one pass
QuickStart is opinionated on purpose: good defaults, less self-inflicted nonsense
If you are explaining setup to new users, start with the wizard — not a giant config dump

⏰

Cron Is a Real Operator Tool

Cron runs inside the Gateway and persists jobs in its shared SQLite state database (legacy ~/.openclaw/cron/ files are migrated on upgrade)
You can target the main session or run isolated cron jobs in their own session
That means reminders, recurring reports, and background chores can be structured instead of jammed into chat history
In plain English: use cron for exact timing, heartbeat for fuzzy recurring check-ins

🛠️

Skills Are Becoming the Real Extension Layer

Skills are AgentSkills-compatible folders with SKILL.md instructions
They load from bundled, managed, or workspace locations with clear precedence
That makes OpenClaw feel less like one bot and more like a serious operating system for agent workflows
If a feature needs an external CLI or API, a skill is often the cleanest path

🚀

Trend worth paying attention to: the best agent stacks are converging on the same recipe — typed native tools, grounded web access, recoverable automation, and explicit setup flows. OpenClaw is leaning hard into that, which is exactly why it feels more usable than the average AI toy.

⚙️ Tool Configuration

Fine-tune tool access with allow/deny lists and provider-specific policies.

🔒 Global Allow/Deny

Prevent specific tools from being sent to model providers.

{
  "tools": {
    "allow": ["group:fs", "browser"],
    "deny": ["exec", "process"]
  }
}

🎯 Provider-Specific

Restrict tools for specific models or providers.

{
  "tools": {
    "profile": "coding",
    "byProvider": {
      "openai/gpt-5.4": {
        "profile": "minimal" 
      }
    }
  }
}

👤 Per-Agent Override

Different tools for different agents.

{
  "agents": {
    "list": [{
      "id": "support-bot",
      "tools": { 
        "profile": "messaging",
        "allow": ["slack"] 
      }
    }]
  }
}

🤔 Tools vs Skills

Understand the difference between built-in tools and installable skills.

🧰

Tools

Built into OpenClaw core
Typed, structured function calls
Always available (can be denied)
Directly integrated with gateway
Examples: browser, exec, cron

🛠️

Skills

Installable from ClawHub
SKILL.md with usage instructions
May require external binaries/APIs
Teach agent to use external tools
Examples: gog, github, weather

💡

Think of it this way: Tools are OpenClaw's native capabilities. Skills teach the agent how to use external CLIs, APIs, and services by providing context and instructions.