OpenClaw Setup Guide:26 Tools + 53 Skills Explained

You’ve installed OpenClaw. Now what?

Tools are scattered across different docs. Skills auto-load by default — some are already active and you don’t even know it. Enable everything and you’re exposed. Disable everything and you’ve wasted the install. Piecing together the full picture from docs and source code takes real effort.

This is my research notes after setting up OpenClaw — compiled from the official docs and GitHub source code, covering what each of the 26 Tools and 53 official bundled Skills does, whether to enable it, how I configured mine, and why. Security analysis is covered in the security guide. This article focuses on what each Tool and Skill does and how to configure them for your needs.

(There are 13,700+ third-party Skills on ClawHub — those are outside the scope of this guide.)

Updated 2026-10-12: This guide now matches the official docs for OpenClaw v2026.10.1 (released 2026-10-10). Four things from the March version will break an older config: image was renamed to view_image; canvas, llm-task, and lobster are now provided by plugins; coding-agent has to be turned on by hand; and approvals.exec no longer decides whether OpenClaw asks before running a command. If you copied my config in March, go straight to “How Do I Configure My Tools and Skills?”.

What’s the Difference Between Tools and Skills?

Many people confuse these two. It’s actually straightforward.

Tools are organs — they determine whether OpenClaw can do something. read and write let it access files, exec lets it run system commands, web_search lets it search like Google, web_fetch lets it read web pages, browser lets it interact with pages (click buttons, fill forms, take screenshots). Without a Tool enabled, it’s like having no hands — nothing gets done.

Skills are textbooks — they teach OpenClaw how to combine Tools to accomplish tasks. gog teaches it how to use Google Workspace for email and calendar, obsidian teaches it how to organize notes, github teaches it how to work with repos, slack teaches it how to send messages to channels. The 53 official Skills cover notes, email, social media, development, smart home, and more.

Does installing a Skill give OpenClaw new permissions? No.

Example: You install the obsidian Skill. OpenClaw now knows how to organize notes — but without the write Tool enabled, it can’t write files at all. Skills are manuals. The real switches are in Tools.

For OpenClaw to actually do something via a Skill, three conditions must be met. Take “read your Gmail” as an example:

1. Configuration: Did you allow OpenClaw to run commands? (Without exec, it can’t even launch a program)
2. Installation: Is the gog bridge tool installed on the machine? (Without it, OpenClaw knows what to do but can’t connect to Google)
3. Authorization: Did you log into your Google account and grant access? (Without authorization, Google won’t let it in)

All three are required. Skills are manuals — whether something actually works depends on these three conditions.

How Are 26 Tools and 53 Skills Organized?

Listing all 26 Tools and 53 Skills flat would be overwhelming. I organize them in concentric circles:

• Layer 1 — Core Capabilities (8 Tools): File access, command execution, web access. Almost everyone enables these.
• Layer 2 — Advanced Capabilities (18 Tools): Browser control, memory, multi-session, automation. Enable as needed.
• Layer 3 — Knowledge Layer (53 Skills): Teaches OpenClaw to work with Google, Obsidian, Slack, etc. Install what you use.

The 26 is the count from when I compiled this in March 2026. As of v2026.10.1, the official docs no longer give a single total for tools. They use a category table instead, split across three layers: tools, skills, and plugins. The tool groups table alone lists 55 tool ids. This guide still walks through the original 26, and the newer ones are collected in “Which Tools Are New in v2026.10.1?” below.

Layer 1: Core Capabilities (8 Tools)

These 8 Tools are OpenClaw’s foundation. With only these enabled, OpenClaw is reactive — you ask, it responds. It can read files, run commands, and search the web, but it won’t remember your preferences across sessions or proactively push you notifications. What turns OpenClaw from a “chatbot” into an “assistant” is Layer 2. But without Layer 1, Layer 2 can’t function.

File Operations: read, write, edit, apply_patch

read is read-only. write and edit can modify files, apply_patch applies code changes. These four are fundamental — most people enable all of them.

Execution & Process Management: exec, process

exec lets OpenClaw run any shell command — install packages, run scripts, manage the system. “Any” is the key word: it can install dependencies for you, but it can also rm -rf your entire machine. Without exec, most tasks fail. With exec and no safeguards, you’ve handed over root access.

That’s why I strongly recommend enabling approval alongside exec — any command that isn’t on the allowlist is shown to you first, and only runs after you confirm. In v2026.10.1 this takes two settings: tools.exec.mode decides whether OpenClaw asks, and approvals.exec forwards the prompt to your chat channel (Telegram, for example):

{
  "tools": {
    "exec": { "mode": "ask" }
  },
  "approvals": {
    "exec": { "enabled": true }
  }
}

approvals.exec on its own does not make OpenClaw ask. On a gateway host, exec runs without asking by default.

Is it annoying? Honestly, yes. But it’s the most basic protection — if the AI ever misjudges or gets hit by a Prompt Injection attack, this gate is your last line of defense.

process manages background processes — list tasks, check output, kill stuck processes. Usually enabled alongside exec.

Web Access: web_search, web_fetch

web_search performs keyword searches, web_fetch reads web page content. Together, they let OpenClaw browse the internet for information.

Layer 2: Advanced Capabilities (18 Tools)

Layer 1 is “can it work at all.” Layer 2 is “does it work well.” These Tools transform OpenClaw from a command executor into a real assistant — one that remembers your preferences, controls a browser, and sends scheduled notifications. But each additional Tool expands the attack surface, so evaluate whether the trade-off is worth it.

Browser: browser, canvas, view_image

browser lets OpenClaw control Chrome — click buttons, fill forms, take screenshots. v2026.3.13 added Chrome DevTools attach mode, letting it connect directly to your signed-in Chrome session without any extensions. I use it for price comparisons, spec research, and adding items to shopping carts. But I always check out myself. The “last mile” involving payments never goes to AI — that’s my line.

canvas is now provided by a plugin (the plugin ships with OpenClaw). It presents widgets and dashboard content on paired macOS panels.

view_image lets OpenClaw “understand” images. It is the tool that used to be called image: any older config that names image has to be renamed, and running openclaw doctor --fix once does it for you.

Memory: memory_search, memory_get

Lets OpenClaw remember information across sessions. After a week of use, it knows I build blogs with Astro, deploy on Azure, and prefer Traditional Chinese — no need to re-explain every time. The longer you use it, the better it knows you.

Multi-Session: sessions series (5 tools)

Run multiple sessions for different tasks simultaneously — one discussing a product idea, another researching travel plans, without interference.

sessions_list and sessions_history view sessions. session_status checks status. sessions_send and sessions_spawn enable inter-session communication and spawning sub-tasks. sessions_yield pauses the main agent to wait for sub-agent results, and subagents manages running sub-agents (check status, stop, etc.)—these two are for multi-agent orchestration.

Messaging: message

Lets OpenClaw send messages to Discord, Slack, Telegram, WhatsApp, iMessage.

I have this enabled but only for sending messages to myself — never for communicating with others on my behalf. The reason is simple: messages sent in your name by AI can’t be unsent. If it misunderstands the context, uses the wrong tone, or gets tricked by Prompt Injection into sending something, you bear the consequences.

I use OpenClaw as the mobile gateway to my entire AI second brain — enabling message lets it proactively push notifications to me: Daily Briefs, task reminders, and to-do alerts, all sent to myself.

Hardware Control: nodes

Cross-device hardware control — remote screenshots, GPS location, camera access.

When I first saw this Tool, I asked myself: when would I need AI to open my camera on its own? I couldn’t think of a scenario. For screenshots, I can just send one via Telegram — one extra step but much more peace of mind. Disabled.

Automation: cron, gateway

cron sets up scheduled tasks. gateway lets OpenClaw restart itself.

Every morning at 6:47, my Telegram receives a Daily Brief prepared by OpenClaw — what I need to do today, pending messages to reply to, and the weather forecast. That’s cron plus message in action, — OpenClaw is the bridge between my entire AI second brain and my phone, so I stay in the loop even away from my computer.

Agent Communication: agents_list

Lists available Agent IDs. OpenClaw supports multi-agent architectures, but official docs don’t go into detail. If you’re running a single OpenClaw instance, you won’t need this.

Voice: tts

tts (text-to-speech) converts OpenClaw’s replies into voice messages — on Telegram, they appear as round voice note bubbles. Supports ElevenLabs, OpenAI TTS, and Edge TTS. Disabled by default; enable auto-voice in config with messages.tts.auto or toggle per session with /tts on.

Plugin Tools: llm-task, lobster

lobster is a workflow engine for defining multi-step processes. llm-task inserts LLM processing steps into workflows.

Both are now optional tools provided by plugins, so neither shows up by default. lobster has to be installed first with openclaw plugins install @openclaw/lobster. llm-task ships with OpenClaw, but you have to set it to enabled: true under plugins.entries. After that, add the tool name to tools.alsoAllow before the agent can call it.

If you’re not using a workflow engine, skip both.

Which Tools Are New in v2026.10.1?

The 26 above are the March list. OpenClaw has added tools since then. The table below follows the categories on the official overview page and only lists tools this guide hasn’t covered yet:

Two more things affect how you configure tools.

Tool profiles. tools.profile picks a base set of tools (minimal, coding, messaging, or full), and tools.allow and tools.deny adjust it from there. To add a few tools on top of a profile, use tools.alsoAllow. You can’t set allow and alsoAllow together. Config validation rejects it.

Tool Search and Code Mode are experimental. With a large catalog, sending every tool definition to the model takes up a lot of room. These two features let the model search first and load a tool only when it needs one. The official docs label both as experimental. OpenClaw’s built-in agent runs use Tool Search automatically when it isn’t configured. Set tools.toolSearch: false to turn it off.

Which of the 53 Official Skills Should You Install?

53 sounds like a lot, but after scanning them you’ll find maybe a dozen are relevant to you. The rest — food delivery, smart home, voice calls — aren’t bad, just irrelevant if they don’t match your use case.

Important: bundled Skills auto-load by default — if the corresponding CLI tool is installed on the system, the Skill activates automatically. It’s not “nothing unless installed” but “everything unless disabled.” To control which Skills are active, use skills.allowBundled in whitelist mode (config example in the “My Config” section below).

There is one exception: coding-agent is now opt-in. Set skills.entries.coding-agent.enabled: true in your config, and make sure a CLI such as claude, codex, or opencode is installed and signed in on the machine.

ClawHub has 13,700+ third-party Skills. Since February 2026, ClawHub has integrated VirusTotal automatic scanning to block malicious Skill downloads, but reviewing before installing is still recommended (see the security guide).

Where Skills Load From: Seven Sources

A Skill with the same name can exist in several places at once. OpenClaw picks which copy to use in this order, highest priority first:

This matters in two ways. First, if you put a Skill in your workspace with the same name as an official one, the official copy is overridden. Second, allowBundled only controls tier 6. The other tiers are not affected by it.

If you run several agents, agents.defaults.skills sets a shared list and agents.entries.<agent>.skills sets a list for one agent. Once an agent has its own list, that list replaces the shared one. It does not merge with it. Setting it to [] means that agent sees no Skills at all.

v2026.10.1 also adds installed-skill search. When you have more Skills than fit in the prompt, the agent can call skills_search to find a relevant one and skills_read to load its full instructions. It only searches Skills you already have. It does not look on ClawHub and it does not install anything.

Organized by use case below.

Notes

4 note-taking Skills: obsidian, notion, apple-notes, bear-notes. Whether they work depends on your deployment.

apple-notes and bear-notes only work on Mac locally — if OpenClaw runs in a VM, they’re out. obsidian operates on local files. I use Obsidian, but my vault is on my Mac while OpenClaw is on an Azure VM, so I handle notes with Claude Code locally rather than through OpenClaw. If you want OpenClaw to manage notes directly and it runs in a VM, notion is cloud-based with no deployment constraints — the path of least resistance.

Productivity

Two email Skills: gog and himalaya. gog integrates the entire Google Workspace (Gmail, Calendar, Tasks, Drive, Docs, Sheets). himalaya uses IMAP/SMTP for sending and receiving only. If you’re on Google, go with gog — more complete, and you can revoke access from your Google account anytime.

Task management has things-mac (Things 3), apple-reminders, and trello. If you already have gog, Google Tasks is included — no extra install needed.

Messaging & Social Media

wacli (WhatsApp), imsg (iMessage), bird (X/Twitter), slack, discord — these Skills give OpenClaw deep access to each platform, including searching message history, syncing conversations, and managing channels. Unlike the message tool (which only sends messages), installing these gives it full access to your data on that platform.

I haven’t installed any of them. The last step in external communication is always manual.

Developer Tools

• github: Operates GitHub via gh CLI, requires OAuth, permissions are controllable
• tmux: Manages multiple terminal sessions
• session-logs: Searches and analyzes past conversation logs
• coding-agent: Calls other AI coding assistants (Claude Code, Cursor, etc.) in the background

I have github, tmux, and session-logs installed. I write code locally with Claude Code, but OpenClaw is always reachable via Telegram — if CI/CD breaks while I’m out, I just ask “check why this PR build failed” on my phone, and it pulls the GitHub Actions error log and tells me the cause.

I haven’t installed coding-agent yet, but the potential is significant — you could install Claude Code on OpenClaw’s VM and let it dispatch coding tasks in the background. Imagine telling OpenClaw via Telegram: “I found this interesting repo on GitHub — clone it, study it, and build a demo site.” It launches Claude Code, executes autonomously, and pushes a notification when done. AI orchestrating AI. I haven’t explored this deeply yet, but it’s on my list. If you try it, turn it on first: in v2026.10.1, coding-agent no longer enables itself.

Password Management

1password lets OpenClaw access your 1Password vault — look up passwords, auto-login, fill forms. Use cases like: “Log me into AWS Console” or “What’s the password for this site?”

But the permission model is all-or-nothing: once authorized, it has access to the entire vault. You can’t limit it to specific entries — whatever you’ve stored, it can read. I chose not to install it. If you really need this, consider creating an “AI-only vault” containing only passwords you’re comfortable sharing with AI.

Other Categories

The categories above are what I actively use or seriously considered. The rest — music playback, smart home, image generation, speech-to-text, food delivery — I haven’t installed. Full list in the appendix.

How Do I Configure My Tools and Skills?

My OpenClaw runs on an Azure VM, operated via Telegram. Paired with Claude Code on desktop, it forms a mobile + desktop dual workflow — mobile for discussions, research, and capturing ideas anytime (conversation history syncs automatically), desktop for execution. I also use it daily for email, calendar, research, and the morning Daily Brief.

Here’s my current config and the reasoning behind each choice.

Tools (22 of 26 enabled)

My rule is simple: if I can’t think of a use case, it stays off.

{
  "tools": {
    "allow": [
      "read", "write", "edit", "apply_patch",
      "exec", "process",
      "web_search", "web_fetch",
      "browser", "view_image",
      "memory_search", "memory_get",
      "sessions_list", "sessions_history", "sessions_send", "sessions_spawn", "session_status",
      "tts",
      "message", "cron", "gateway", "agents_list"
    ],
    "deny": ["nodes", "canvas"],
    "exec": { "mode": "ask" }
  },
  "approvals": {
    "exec": { "enabled": true }
  }
}

22 enabled, 2 explicitly denied: nodes (can’t think of a scenario) and canvas (don’t need it). llm-task and lobster are plugins now. If you haven’t installed or enabled them they don’t exist, so they no longer belong in deny. sessions_yield and subagents are for multi-agent orchestration — not using them yet so they’re not listed. exec has approval enabled. message only sends to myself.

Skills (9 of 53 enabled)

As mentioned earlier, bundled Skills auto-load by default. I use allowBundled whitelist to restrict to only what I need:

{
  "skills": {
    "allowBundled": [
      "gog", "github", "tmux", "session-logs",
      "weather", "summarize", "clawhub",
      "healthcheck", "skill-creator"
    ]
  }
}

In short: gog for email and calendar, github for repos, and the rest are utilities for Daily Brief and system management.

How to Automate Tasks with Your AI Agent

This is where OpenClaw stops being a chatbot and starts being infrastructure. The combination of cron (scheduling) and message (push notifications) turns it into an automation engine that works while you sleep.

The pattern is always the same: trigger + action + deliver. Define when it runs, what it does, and where results go. Here are the automations I actually use:

Daily Brief — Every morning at 6:47, my Telegram receives a briefing: today’s calendar, pending emails that need replies, weather forecast, and any CI/CD failures overnight. This single automation replaced my habit of checking five different apps before coffee.

Email Triage — Twice a day, OpenClaw scans my inbox, categorizes messages by urgency, and sends me a summary. Newsletters get archived. Anything requiring action gets flagged with a one-line summary. I went from 30 minutes of inbox management to 5.

CI/CD Monitoring — When a GitHub Actions workflow fails, OpenClaw reads the error log, identifies the likely cause, and pushes a Telegram message with the diagnosis. I’ve fixed production issues from my phone while standing in line for coffee.

Content Research — Daily, OpenClaw collects trending discussions from specific subreddits, Hacker News threads, and RSS feeds I follow, then compiles a digest of potential writing topics. It doesn’t write for me — it surfaces what’s worth writing about.

The setup isn’t complicated. Each automation is a cron entry that triggers a prompt, and the prompt tells OpenClaw which tools to use and where to send results. The hard part isn’t the config — it’s figuring out which parts of your routine are worth automating. Start with one that saves you the most daily friction, get it working, then add more.

How Do You Get Started?

You don’t need all 26 Tools. The 53 bundled Skills default to all-on (except coding-agent) — use allowBundled to keep only what you need. Open your openclaw.json and start with three principles:

1. If you can’t think of a use case, leave it off
2. More capability, more control — enable approval for exec, only message yourself
3. The last mile is always manual — checkout, sending messages, posting publicly — anything irreversible stays with you

My config above works as a starting point. Copy it, then trim to fit your needs. For security settings, read it alongside the security guide. For deployment costs, see the deploy cost guide.

Appendix: Full Reference

Last updated: 2026-10-12