OpenClaw for beginners: set up and complete your first task

OpenClaw becomes useful when it produces something you can inspect. In this guide, you will install the agent, connect a model, give it access to a bounded workspace, and ask it to create a real meeting checklist. Then you will read the file back and verify the result.

This is part two of Hammer's Personal AI Assistant series. If agents, models, and tools are new to you, start with the Hermes Agent guide to personal AI assistants. This article focuses on OpenClaw. The comparison between the two tools comes in part three.

Who this OpenClaw guide is for

This guide is for anyone who wants to test a personal AI agent on their own computer or server and get a result they can review. You need to be comfortable opening a terminal and following an installation guide, but you do not need to be a programmer.

In 15–20 minutes, you will have:

• an installed OpenClaw client
• a verified model connection
• a local Gateway and Control UI
• a meeting checklist with exactly five items
• a receipt with the path, item count, and first item

We will not cover migration from another agent, advanced plugins, multi-agent systems, or a broad tour of use cases.

Before you start

You need macOS, Linux, WSL2, or Windows, plus access to an AI model. OpenClaw's recommended installer handles Node when it is missing. If you already manage Node yourself, you can use npm instead; OpenClaw currently requires Node 22.22.3+, 24.15+, or 25.9+, and recommends Node 26.

Choose an empty or clearly bounded folder for the test. Do not use a customer project or a folder containing passwords, private keys, or personal data. We will make a real file change, but only inside a disposable test area.

Source: OpenClaw Install.

1. Install OpenClaw and verify the model

On macOS, Linux, or WSL2, use the recommended route:

curl -fsSL https://openclaw.ai/install.sh | bash

On Windows, use the PowerShell command from the OpenClaw installation page. The installer normally starts onboarding for you. If the package is already installed, run:

openclaw onboard

Onboarding now starts with model access. It looks for usable logins and API keys, runs a real model completion, and saves only a route that works. Choose the provider and sign-in method you intend to use. If the model test fails, fix the login instead of continuing with an agent that only appears ready.

Once the model works, OpenClaw helps you configure the workspace, Gateway, and optional channels. The full classic wizard remains available through openclaw onboard --classic when you need to select every setting yourself.

Source: OpenClaw CLI onboarding.

2. Understand the Gateway and open Control UI

The Gateway is OpenClaw's always-on process. It connects agent runs, Control UI, APIs, and chat channels. A standard local setup normally uses loopback, port 18789, and token authentication.

Check that it is running:

openclaw gateway status

Then open Control UI at:

http://127.0.0.1:18789/

A new browser may need both the Gateway token and a one-time device approval. Loopback keeps the interface on the same computer, but it does not remove authentication.

In our isolated test, Gateway health returned ok, Control UI returned HTTP 200, and the OpenAI model probe passed before we ran the file task.

Sources: OpenClaw Gateway. OpenClaw Control UI.

3. Limit the first task to file tools

This exercise only needs file reads and writes. We limited the tool policy to the file group and restricted file tools to the agent workspace:

openclaw config set tools.allow '["group:fs"]' --strict-json
openclaw config set tools.fs.workspaceOnly true --strict-json

Start a new session after changing the policy, then test the boundary yourself. OpenClaw can use different agent runtimes for different model routes. Our passing acceptance test used the explicit OpenClaw runtime: the file tool rejected ../outside-boundary-test.md, while reads and writes inside the workspace succeeded.

A technical policy does not replace a sensible workspace. Keep the first task in a dedicated demo folder and give the agent a precise change boundary.

Sources: OpenClaw model providers and agent runtimes. OpenClaw security.

4. Run the first useful task

Open a fresh OpenClaw chat and paste:

Work only inside your OpenClaw workspace.

Create the first-task-demo folder if it does not exist, then create
first-task-demo/meeting-checklist.md with a clear heading and exactly
five unchecked Markdown boxes for preparing a work meeting.

Use the file tools, not shell commands. Read the file back and return
a receipt containing:
1. the resolved absolute file path
2. the checklist item count
3. the first checklist item verbatim

Change nothing outside first-task-demo.

The prompt gives the agent a concrete job, a fixed location, a measurable result, and a requirement to read the file back. It does not merely ask the model to say the work is complete.

5. Check the receipt, not just the chat response

In Hammer's isolated run with OpenClaw 2026.7.1-2, the agent used the write and read tools without tool errors. The result was:

• file: …/.openclaw/workspace/first-task-demo/meeting-checklist.md
• checklist items: 5
• first item: - [ ] Confirm agenda and desired outcome
• other files in first-task-demo: none

We also inspected the file outside the agent response. It existed, contained exactly five unchecked boxes, and the demo folder contained only meeting-checklist.md.

This is the sort of receipt worth carrying into real workflows: the correct location, an exact count, a concrete first item, and an external check. If the result does not match, ask the agent to explain the difference before allowing more work.

6. Add Telegram only after the local path works

OpenClaw has a production-ready Telegram channel for direct messages and groups. Create a bot with the real @BotFather, store its token as a secret, and configure the channel through OpenClaw's channel setup or masked channel wizard. Telegram does not use openclaw channels login telegram.

Once the Gateway is running, you can approve the first direct message with:

openclaw pairing list telegram
openclaw pairing approve telegram <CODE>

The pairing code expires after one hour. For an owner-operated bot, an explicit numeric user allowlist is clearer than an open channel. Never put the bot token in a prompt or an ordinary chat transcript.

Telegram does not make the agent cloud-hosted by itself. The Gateway and the computer or server hosting it must remain online. If the host shuts down, the channel stops responding until the Gateway returns.

We verified this Telegram flow against the current official documentation, but did not use Adam's existing bot or live channel in the test.

Source: OpenClaw Telegram setup and pairing.

Common mistakes in a first OpenClaw test

• The model was never verified. Run onboarding again and require a passing completion test.
• The Gateway is not running. Check openclaw gateway status before debugging the chat.
• The workspace is too broad. Start in a dedicated demo folder and request one named file.
• The policy is assumed to work. Run a boundary test with your chosen model and agent runtime.
• Telegram is expected to work while the host is off. Put the Gateway on a host that is genuinely meant to stay online if you need an always-available channel.

Next step: turn the test into a reliable routine

Run the task again with your own meeting template. Add an approval stop before future file changes and save the receipt beside the result. Once that works, connect Telegram and ask the agent to read the checklist back from your phone.

If you want to make a personal agent useful inside a real team, Hammer Automation can help map the workspace, secrets, scoped permissions, approvals, and run receipts. Start with one real workflow and require the same clear receipt before the agent receives broader access.

How this guide was made

A running Hermes agent performed the source review, isolated OpenClaw installation, model and Gateway checks, Swedish and English file tasks, boundary test, writing, media production, and publishing workflow. Adam and Hammer review the live output and improve the automation over time. Human oversight remains part of the process even when the agent does most of the work.