OpenClaw

OpenClaw: Advanced features

This is the seventh of eight OpenClaw pages. It covers five power features for after your first week, each with what it is for, how OpenClaw's docs turn it on, and what it trades away.

These steps match OpenClaw 2026.11.3 and its docs as of 14 November 2026. None of them is needed for the jobs on the earlier pages, so take them one at a time, and keep the approval switch on ask and the file tools inside the workspace throughout.

A real sandbox with Docker

What it is for. The sandbox moves the agent's commands off your real computer and into a container. A mistake or a prompt injection then lands there instead of on your files. It is the step that makes longer or unattended jobs reasonable.

How to turn it on. Install Docker Desktop and start it first, because Docker is OpenClaw's default sandbox backend. Then paste these lines one at a time.

openclaw config set agents.defaults.sandbox.mode all
openclaw doctor --fix
openclaw gateway restart

The docs say to set the mode to all or non-main and then run openclaw doctor --fix, which offers to build the sandbox image, openclaw-sandbox:trixie-slim. Accept that build. Use all, because non-main leaves your main chat on your real computer. To see what the sandbox is actually doing, run openclaw sandbox explain.

Once it is on, the defaults are strict. Commands get no internet, and the sandbox works in its own folder under ~/.openclaw/sandboxes instead of your workspace. The setting agents.defaults.sandbox.workspaceAccess changes that: none is the default, ro lets the sandbox read your workspace without changing it, and rw lets it change your workspace files.

Commands and the file tools move into the sandbox, but the Gateway itself, add-ons, and MCP servers stay on your computer. This guide could not confirm that the Docker sandbox works with the native Windows install from the PowerShell path. On Windows, check openclaw sandbox explain before you rely on it.

Your own skills

What it is for. A skill is a saved set of instructions the agent loads when it needs them, so you stop repeating the same request. OpenClaw reads skills from the skills folder in your workspace, and each skill is a folder with one SKILL.md file. The shared Skills and MCP page compares skills and MCP across every agent in this guide and explains how to choose safe ones.

How to turn it on. Make a folder for the skill. On a Mac, paste this into Terminal.

mkdir -p ~/.openclaw/workspace/skills/summary-style

On Windows, paste this into PowerShell.

New-Item -ItemType Directory -Force "$env:USERPROFILE\.openclaw\workspace\skills\summary-style"

In that folder, save a plain text file named SKILL.md with the text below inside it.

---
name: summary-style
description: Write summaries as four short lines with the most useful fact first.
---

# Summary style

When the user asks for a summary, write four short lines.
Put the most useful fact first, and skip any greeting.
Do not run commands, and do not open any website.

The name uses lowercase letters, digits, and hyphens, and it should match the folder name. The description stays on one line, under 160 characters. Check that OpenClaw found it with openclaw skills list, type /new in the chat, and then use /skill summary-style to call it by name.

The agent can also write skills for itself, through a feature called Skill Workshop. Those skills apply at once, but each change keeps the previous version, so you can review them with openclaw skills workshop list and openclaw skills workshop changes, and hide one with openclaw skills workshop archive <name>.

Skills from ClawHub are someone else's code. OpenClaw's docs say to treat third-party skills as untrusted code and read them before you turn them on, and openclaw skills verify @owner/<slug> shows ClawHub's check of a skill.

Tools from MCP servers

What it is for. An MCP server is a program or online service that lends the agent extra tools. One might search a particular website, and another might read a calendar. OpenClaw connects to the server and adds its tools to the agent's list.

How to turn it on. In the chat page, open Settings → MCP and select Add server under Configured servers. Give it a name, and pick a transport: Streamable HTTP, SSE, or Stdio. For the first two you enter the server's web address. For Stdio you enter a command, which means OpenClaw starts that program on your computer. Then check that it answers.

openclaw mcp doctor <name> --probe

Saving a server does not prove it works, and the docs say this check is the proof. You can limit a server to some of its tools with the --include option of openclaw mcp add, or the tool filter in Settings. The docs also say a server's tools go through the same tool rules as everything else, so connecting one does not get around your settings. Keep passwords and keys out of the server's settings, and use OpenClaw's secret storage instead.

Several agents at once

What it is for. One Gateway can run several separate agents, each with its own workspace, its own settings, and its own chat history. That lets you keep a cautious agent apart from one you trust with more.

How to turn it on. Add an agent by name:

openclaw agents add work

Then check the list with openclaw agents list --bindings. In the chat page, the Agents page shows each agent, and Manage agents changes them.

Each agent can have its own sandbox and tool limits, set under agents.entries.*.sandbox and agents.entries.*.tools. The docs' example gives a family agent a sandbox mode of all and lets it use only the read tool, denying exec, write, edit, and apply_patch. For commands, each agent's workspace is still only a starting place, not a wall, unless its sandbox is on. An agent's own tools.fs.workspaceOnly setting overrides the one from page 2, so never set it to false for a new agent.

A local model with Ollama

What it is for. A local model runs on your own computer, so it costs nothing per use and your prompts stay on your machine.

How to turn it on. Install Ollama from ollama.com/download, then download a model:

ollama pull gemma4

Run openclaw onboard, choose Ollama, and then choose Local only. Setup prefers a model that supports tools and has at least 16K of context. Check that OpenClaw sees it, then make it the main model:

openclaw models list --provider ollama
openclaw models set ollama/gemma4

After setup, run openclaw exec-policy show again to confirm the approval switch still says ask, and openclaw config get tools.fs.workspaceOnly to confirm the file tools still say true.

OpenClaw's docs warn that smaller or heavily compressed models are easier to trick with a prompt injection. They recommend a large model for any agent with tools, and loop detection for smaller models, which Fixes shows how to turn on.

Shared advanced topics

These pages apply to every agent, and these are the ones that fit OpenClaw:

Sources

All checked 14 November 2026.

Next

Continue to Persona and rules.

Checked 14 November 2026 · prices, limits, and versions change