SMF Clearinghouse

What you'll get

A self-hosted OpenClaw agent running on your Linux machine. It will connect to a local Ollama model and run a simple tool so you can verify the full loop: prompt → model → tool → result → reply.

Prerequisites

Ubuntu 24.04 or another Linux distribution
Ollama running locally with at least a 9B parameter model (see the Ollama CUDA recipe)
Node.js 22+ and npm
Git

Step 1: Install the OpenClaw CLI

npm install -g openclaw
openclaw --version

If you prefer not to install globally, use npx openclaw instead of openclaw in the commands below.

Step 2: Create an agent workspace

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

This creates:

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

Step 3: Configure the agent

Edit openclaw.json:

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

Make sure the model name matches what Ollama serves:

ollama list

Step 4: Add a simple tool

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" }
  },
  required: ["city"]
};

export async function run({ city }) {
  // For this recipe, return a mock result. In production, call wttr.in or a weather API.
  return {
    city,
    condition: "sunny",
    temperature_c: 22
  };
}

Step 5: Run the agent

openclaw run "What is the weather in Boston?"

Expected flow:

OpenClaw sends your prompt to qwen3.5:9b via Ollama.
The model decides it needs weather data.
OpenClaw calls get_weather({ city: "Boston" }).
The result is returned to the model.
The model answers: "The weather in Boston is sunny and 22°C."

Step 6: Verify the loop

Run with verbose logging:

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

Step 7: Add memory

By default, OpenClaw stores conversation context in memory/. Test it:

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

The second response should recall "Helix."

Troubleshooting

Symptom	Fix
`model not found`	Run `ollama pull qwen3.5:9b` or update `openclaw.json` to match `ollama list`.
Tool is never called	Check the tool's `description` is specific. Add `required` to parameters.
High latency	Use a smaller model or enable GPU offload (see Ollama CUDA recipe).
Port connection error	Confirm Ollama is listening on `http://localhost:11434` with `curl http://localhost:11434/api/tags`.

Next step

Connect the agent to a real tool such as file search, GitHub issues, or a database. See the OpenClaw skills documentation and the Cline + local model recipe.

Build Your First OpenClaw Agent