Installation Guide

OpenClaw
+ llama.cpp

Install OpenClaw from scratch on Ubuntu and connect it to a local llama.cpp server. No API costs. No cloud. Full inference on your own GPU.

Model Qwen3.5-35B-A3B Q8_0

GPU NVIDIA RTX A6000

Port 8080 (llama.cpp)

Node v22+

Cost $0 / token

Prerequisites

✓

llama.cpp server already running

Your llama.cpp server must be running and accessible before starting. Verify it is up:

bash · verify llama.cpp

curl http://localhost:8080/v1/models
# Should return model list JSON

Endpoint

localhost:8080

Model ID

Qwen3.5-35B-A3B

Context

32768 tokens

Speed

~77 tok/sec

Installation Steps

Install Node.js 22

OpenClaw requires Node.js v22 or higher. Install via NodeSource.

bash

curl -fsSL https://deb.nodesource.com/setup_22.x | sudo -E bash -
sudo apt-get install -y nodejs
node --version   # should print v22.x.x

Install OpenClaw

Install the latest OpenClaw globally via npm.

bash

npm install -g openclaw@latest
openclaw --version

Run Onboarding Wizard

Run the wizard to initialize the workspace and install the gateway daemon. Skip provider selection — we configure llama.cpp manually in the next step.

bash

openclaw onboard --install-daemon

ℹ Complete the wizard with any provider selection. The config file will be overwritten in the next step with the correct llama.cpp settings.

Configure for llama.cpp

Replace the contents of ~/.openclaw/openclaw.json with the following config.

json · ~/.openclaw/openclaw.json

{
  "models": {
    "providers": {
      "llamacpp": {
        "baseUrl": "http://127.0.0.1:8080/v1",
        "apiKey": "dummy",
        "api": "openai-completions",
        "models": [
          {
            "id": "Qwen3.5-35B-A3B",
            "name": "Qwen3.5-35B-A3B",
            "reasoning": true,
            "input": ["text"],
            "cost": {
              "input": 0, "output": 0,
              "cacheRead": 0, "cacheWrite": 0
            },
            "contextWindow": 32768,
            "maxTokens": 8192
          }
        ]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "llamacpp/Qwen3.5-35B-A3B"
      },
      "workspace": "/home/Ubuntu/.openclaw/workspace"
    }
  },
  "gateway": {
    "port": 18789,
    "mode": "local",
    "bind": "loopback",
    "auth": {
      "mode": "token"
    }
  }
}

openai-completions ✓ USE THIS

Standard OpenAI format. Correct for llama.cpp, vLLM, LiteLLM, and Ollama. Returns standard chat/completions response.

openai-responses NOT FOR LLAMA.CPP

Only for specialized APIs like MiniMax that return non-standard response formats. Do not use with llama.cpp.

Start the Gateway

Start OpenClaw gateway in verbose mode to confirm llama.cpp connection.

bash

openclaw gateway --port 18789 --verbose

Test the Agent

Send a message directly to the agent and confirm it is routing through llama.cpp.

bash

openclaw agent --message "What is the derivative of x squared times sin 3x?"

# Check gateway status
openclaw status

# Run full diagnostics
openclaw doctor

✓ If the response arrives and matches the quality you saw from the Python test, llama.cpp routing is working correctly. You can now add any channel (Telegram, Discord, Slack) on top of this config.

Add Channels (Optional)

To add Telegram or any other channel, extend the config with the channels section. The llama.cpp provider config stays exactly the same.

json · add to openclaw.json

"channels": {
  "telegram": {
    "botToken": "YOUR_BOT_TOKEN_HERE",
    "dmPolicy": "pairing"
  }
}

⚠ After adding a channel, restart the gateway: openclaw gateway restart

OpenClaw + llama.cpp

OpenClaw
+ llama.cpp