Ollama Local Model

ollama show gemma4:e4b-128k --modelfile | grep num_ctx
# Should show: PARAMETER num_ctx 131072

2. Add provider in openclaw.json

Add or update the ollama-local provider under models.providers. Use openai-completions API (NOT native ollama API) to avoid context window probe issues:

{
  "models": {
    "providers": {
      "ollama-local": {
        "baseUrl": "http://localhost:11434/v1",
        "api": "openai-completions",
        "apiKey": "ollama-local",
        "models": [
          {
            "id": "gemma4:e4b-128k",
            "name": "Gemma 4 E4B 128k (Local)",
            "reasoning": false,
            "input": ["text"],
            "cost": { "input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0 },
            "contextWindow": 131072,
            "maxTokens": 8192
          }
        ]
      }
    }
  }
}

Critical notes:

• baseUrl must be http://localhost:11434/v1 (with /v1) — this is Ollama's OpenAI-compatible endpoint
• api must be openai-completions — the native "ollama" API has a bug where it probes context window and always reports 8192, regardless of the model's actual num_ctx
• contextWindow must be >= 16000 (OpenClaw's hard minimum) and >= 32000 recommended for tool use
• id must exactly match the Ollama model tag (e.g., gemma4:e4b-128k)

3. Add model to agent defaults

Under agents.defaults.models, add the model so it appears in the model picker:

{
  "agents": {
    "defaults": {
      "models": {
        "ollama-local/gemma4:e4b-128k": {
          "alias": "Gemma4-Local-128k"
        }
      }
    }
  }
}

4. Add auth profile

Add a dummy API key to the auth profiles store. Ollama doesn't need real auth but OpenClaw's openai-completions adapter requires an entry:

Edit ~/.openclaw/agents/main/agent/auth-profiles.json and add:

{
  "ollama-local:default": {
    "type": "api_key",
    "provider": "ollama-local",
    "key": "ollama"
  }
}

5. Verify

After saving, OpenClaw will auto-detect the config change and restart. Check logs:

tail -10 ~/.openclaw/logs/gateway.log
# Should show: config change detected; evaluating reload
# Then: ready (X plugins...)

# If errors occur:
grep "ollama-local" ~/.openclaw/logs/gateway.err.log | tail -10

Troubleshooting

"Model context window too small (8192 tokens)"

• You're using "api": "ollama" instead of "api": "openai-completions"
• Or your contextWindow in the config is set below 16000
• Fix: Switch to openai-completions API with /v1 endpoint

"No API key found for provider ollama-local"

• Missing auth profile entry in ~/.openclaw/agents/main/agent/auth-profiles.json
• Fix: Add the dummy ollama-local:default entry as shown in Step 4

"Something went wrong" (generic error)

• Check both log files:

  tail -20 ~/.openclaw/logs/gateway.err.log
  tail -20 ~/.openclaw/logs/gateway.log
  

• Common causes: context too small, missing auth, model not running in Ollama

Model not appearing in picker

• Ensure model is listed in both models.providers.ollama-local.models[] AND agents.defaults.models
• The provider key format is ollama-local/<model-id> (e.g., ollama-local/gemma4:e4b-128k)

RAM Considerations

Large context windows consume significant RAM/VRAM:

• 8k context: ~10GB for 8B model
• 32k context: ~12GB for 8B model
• 128k context: ~20GB+ for 8B model

If you hit OOM, reduce num_ctx in the Ollama Modelfile and contextWindow in openclaw.json accordingly. Minimum usable is 32768.