# LLM Configuration and Model Strategy

Direct API, unified routing via OpenRouter, or fully local inference with Ollama — choose based on cost, privacy, and capability.

25 minutes

3 provider options

## Choosing a Model Strategy

**Maximum capability.** Use Claude or GPT-4o via direct API or OpenRouter. Best for complex instructions and prompt injection resistance.

**Cost efficiency at volume.** Use OpenRouter with model failover. Route simple tasks to cheap models, complex tasks to heavy ones.

**Full data privacy.** Use Ollama locally. No API calls leave your VPS. Tradeoff: less capable and needs more RAM (8 GB+ for 7B models).

**Flexibility.** Use OpenRouter. One API key, one billing dashboard, 300+ models. Swap models without reconfiguring OpenClaw.

## Option A: Direct Anthropic Connection

The simplest configuration and the recommended starting point. Sign up at console.anthropic.com and set a monthly spending limit immediately.

### Configure Anthropic

```
openclaw config set agent.provider anthropic
openclaw config set agent.api_key "sk-ant-your-key-here"
openclaw config set agent.model claude-sonnet-4-5
```

### Available Models

| Model | Best For | Cost |
| --- | --- | --- |
| claude-opus-4-5 | Complex reasoning, autonomous task chains | High |
| claude-sonnet-4-5 | Everyday tasks, balanced speed and capability | Medium |
| claude-haiku-4-5 | High-volume simple tasks, fast responses | Low |

### Verify

```
openclaw config verify
openclaw gateway restart
```

## Option B: OpenRouter as a Unified Layer

OpenRouter accepts a single API key and routes requests to the model you specify. One billing account, automatic failover, and access to 300+ models.

### Configure OpenRouter

```
openclaw config set agent.provider openrouter
openclaw config set agent.api_key "sk-or-your-key-here"
openclaw config set agent.model anthropic/claude-sonnet-4-5
```

### Configure Model Failover

```
openclaw config set agent.model anthropic/claude-sonnet-4-5
openclaw config set agent.fallback_model openai/gpt-4o
openclaw config set agent.fallback_on_error true
```

### Route Different Agents to Different Models

```
# High-capability agent for complex work
openclaw agents update work-agent --model anthropic/claude-opus-4-5

# Lightweight agent for simple tasks
openclaw agents update reminder-agent --model google/gemini-flash-1.5
```

## Option C: Ollama for Local Inference

Ollama serves open-weight models locally. No API costs, no data leaving the server.

### System Requirements

| Model Size | Min RAM | Recommended |
| --- | --- | --- |
| 7B params | 8 GB | 12 GB |
| 13B params | 16 GB | 20 GB |
| 33B params | 32 GB | 48 GB |
| 70B params | 64 GB | 80 GB |

### Install Ollama

```
curl -fsSL https://ollama.com/install.sh | sh

# Pull a model
ollama pull llama3.2:3b       # 2 GB, fits in 4 GB RAM
ollama pull qwen2.5-coder:7b  # 4.7 GB, fits in 8 GB RAM

# Test the model
ollama run llama3.2:3b "Hello, are you working?"
```

### Secure Ollama

Restrict Ollama to localhost:

```
sudo systemctl edit ollama

# Add to the [Service] section:
[Service]
Environment="OLLAMA_HOST=127.0.0.1:11434"

sudo systemctl daemon-reload
sudo systemctl restart ollama
```

### Configure OpenClaw for Ollama

```
openclaw config set agent.provider ollama
openclaw config set agent.base_url "http://127.0.0.1:11434"
openclaw config set agent.model qwen2.5-coder:7b

openclaw config verify
```

### Ollama on a Separate GPU Server

For faster inference, run Ollama on a GPU server (Vast.ai, Lambda Labs, Hetzner GPU) and point your VPS at it:

```
# On the GPU server
export OLLAMA_HOST=0.0.0.0:11434
ollama serve

# On your OpenClaw VPS
openclaw config set agent.provider ollama
openclaw config set agent.base_url "http://gpu-server-ip:11434"
openclaw config set agent.model llama3.3:70b
```

Add your VPS IP to the GPU server's firewall. Never expose Ollama to the public internet without authentication.

## Session Pruning and Context Management

### Context configuration

```
# Maximum context tokens before pruning
openclaw config set agent.max_context_tokens 100000

# Strategy: 'sliding' keeps recent, 'summarize' compresses old context
openclaw config set agent.context_strategy sliding
```

The `summarize` strategy is more expensive but produces more coherent long-running memory.

## Model Verification Checklist

### Verify everything

```
# Run doctor for overall health
openclaw doctor

# Verify model connection
openclaw config verify

# Check channel-agent assignments
openclaw channels list
openclaw agents list
```

Every channel should show connected, and every channel should have an agent assigned. Send a test message on each channel and confirm responses arrive.