Connect OpenClaw to Ollama: Local AI Agent in Practice

After installing both Ollama and OpenClaw, the final step is wiring them together correctly. This updated guide uses the same practical approach as the Chinese version: edit ~/.openclaw/openclaw.config and point OpenClaw to your Ollama host.

Target Architecture

plain text

OpenClaw
  -> Ollama API (192.168.10.20:11434)
  -> qwen2.5:14b
  -> response

1) Configure Models in `~/.openclaw/openclaw.config`

json

models: {
    mode: 'merge',
    providers: {
      ollama: {
        baseUrl: 'http://192.168.10.20:11434/v1',
        apiKey: 'ollama-local',
        api: 'openai-completions',
        models: [
          {
            id: 'qwen2.5:14b',
            name: 'Qwen-14b',
            api: 'openai-completions',
            reasoning: false,
            input: [
              'text',
            ],
            cost: {
              input: 0,
              output: 0,
              cacheRead: 0,
              cacheWrite: 0,
            },
            contextWindow: 32768,
            maxTokens: 32768,
          },
        ],
      },
      'custom-192-168-10-20-11434': {
        baseUrl: 'http://192.168.10.20:11434/v1',
        apiKey: 'ollama-local',
        api: 'openai-completions',
        models: [
          {
            id: 'qwen2.5:14b',
            name: 'qwen2.5:14b (Custom Provider)',
            reasoning: false,
            input: [
              'text',
            ],
            cost: {
              input: 0,
              output: 0,
              cacheRead: 0,
              cacheWrite: 0,
            },
            contextWindow: 16000,
            maxTokens: 4096,
          },
        ],
      },
    },
  },

The key is to keep baseUrl and model IDs aligned with your actual Ollama endpoint and installed model.

2) Configure Agents

json

agents: {
    defaults: {
      model: {
        primary: 'custom-192-168-10-20-11434/qwen2.5:14b',
      },
      models: {
        'openai-codex/gpt-5.4': {},
        'custom-192-168-10-20-11434/qwen2.5:14b': {},
      },
      workspace: '/root/.openclaw/workspace',
      compaction: {
        mode: 'safeguard',
      },
      maxConcurrent: 4,
      subagents: {
        maxConcurrent: 8,
      },
    },
  },

Keep defaults first, verify stability, then tune prompts and tools.

3) End-to-End Validation

Use a simple prompt:

plain text

Explain the benefits and limitations of running LLMs locally.

A normal model response means your chain is working:

plain text

OpenClaw -> Ollama API -> qwen2.5:14b -> Response

Troubleshooting Checklist

Connection error: ensure ollama serve is running and the configured host (for example 192.168.10.20:11434) is reachable.
Model missing: run ollama list, then ollama pull qwen2.5:14b if needed.
Slow response: check local CPU/RAM pressure or switch to a lighter model.

Wrap-up

You now have a complete local loop from model serving to agent orchestration. Next, you can extend this setup with tool calls (web_fetch, web_search) and proxy-enabled automation workflows.

Connect OpenClaw to Ollama: Local AI Agent in Practice

Key Takeaways