Build Your First OpenClaw Agent

The fastest path is the install script. It handles Node, Git, and the workspace layout for you.

# Linux / macOS / WSL2
curl -fsSL https://openclaw.ai/install.sh | bash

On Windows PowerShell:

iwr -useb https://openclaw.ai/install.ps1 | iex

Verify the install:

openclaw --version

You should see a version number, not a command-not-found error.

openclaw onboard

The wizard asks:

• Local Gateway or remote host
• Model provider (choose Ollama for this recipe)
• Channels you want enabled (skip them for now — you can add Telegram, Discord, Slack later)
• Workspace location

For the fastest first chat without configuring channels:

openclaw dashboard

This opens the Control UI in your browser at http://127.0.0.1:18789/.

OpenClaw keeps each agent in its own workspace. Create one for this recipe:

mkdir -p ~/.openclaw/agents/first-agent
cd ~/.openclaw/agents/first-agent
openclaw init

You should see a scaffold like this:

first-agent/
  openclaw.json      # agent configuration
  tools/             # custom tool definitions
  memory/            # durable memory storage
  prompts/           # prompt templates

Edit openclaw.json:

{
  "agent": {
    "name": "first-agent",
    "model": "ollama/qwen3.5:9b",
    "systemPrompt": "You are a helpful local assistant. Use the available tools when needed. Keep answers concise."
  },
  "tools": {
    "registry": ["./tools"]
  },
  "memory": {
    "store": "./memory"
  }
}

Make sure the model matches what Ollama serves on your machine:

ollama list

If you only pulled a different model, update the model field accordingly. Good starter models include qwen3.5:9b, llama3.2:3b, or gemma2:9b.

Tools are what turn a chat model into an agent. Create tools/weather.js:

export const name = "get_weather";
export const description = "Get the current weather for a city.";
export const parameters = {
  type: "object",
  properties: {
    city: { type: "string", description: "City name, e.g. Boston" }
  },
  required: ["city"]
};

export async function run({ city }) {
  // This recipe returns mock data so you can verify the loop.
  // In production, replace this with a real call to wttr.in or a weather API.
  return {
    city,
    condition: "sunny",
    temperature_c: 22,
    source: "mock"
  };
}

Save it. OpenClaw automatically discovers any .js or .ts file in tools/.

openclaw run "What is the weather in Boston?"

Expected flow:

1. OpenClaw sends your prompt to the local Ollama model.
2. The model recognizes it needs weather data.
3. OpenClaw calls get_weather({ city: "Boston" }).
4. The tool result is returned to the model.
5. The model answers: "The weather in Boston is sunny and 22°C."

If you see that answer, the agent loop is alive.

Memory lets the agent recall things across separate conversations. OpenClaw stores session context in the memory/ folder by default. Test it:

openclaw run "Remember that my favorite editor is Helix."
openclaw run "What is my favorite editor?"

The second response should recall "Helix." If it does, memory is working.

When something goes wrong, visibility is everything:

OPENCLAW_LOG_LEVEL=debug openclaw run "What is 7 * 13?"

You should see:

• The raw model request
• Tool selection reasoning
• Tool output
• Final formatted answer

Once the local loop works, add a messaging channel so you can talk to the same agent from anywhere:

openclaw channels add telegram
openclaw channels add discord

Each channel has its own setup prompts. For a private test, the web dashboard at http://127.0.0.1:18789/ is the safest place to start.