DEV Community

Jenny Met
Jenny Met

Posted on

OpenClaw Custom Model Configuration: Connect Any AI Model You Want

#openclaw #ai #llm #devops

One of OpenClaw's biggest advantages as an open-source AI coding assistant: it doesn't lock you into any single model vendor. Through its flexible Provider system, you can freely connect to OpenAI, Anthropic, Google, DeepSeek, and other cloud models, run local models via Ollama or llama.cpp, or use an API gateway to access 627+ models with one key.

This is the complete guide to configuring custom models in OpenClaw.

Understanding the Provider System

Every model in OpenClaw is referenced in provider/model format — like openai/gpt-5.1-codex or anthropic/claude-opus-4-20250918. The Provider determines the API endpoint, auth method, and request format.

Built-in Providers

Provider Description Typical Models
openai OpenAI official API gpt-5.1, gpt-5.1-mini
anthropic Anthropic official API claude-opus-4, claude-sonnet-4
google Google AI Studio gemini-2.5-pro, gemini-2.5-flash
google-vertex Google Cloud Vertex AI gemini-2.5-pro (enterprise)
openai-codex OpenAI Codex codex-mini-latest
opencode OpenCode-compatible endpoints community models

Quick CLI check:

# List all configured models
openclaw models list

# Switch primary model
openclaw models set anthropic/claude-opus-4-20250918
Enter fullscreen mode Exit fullscreen mode

API Key Configuration & Rotation

OpenClaw supports multi-key rotation with this priority order:

# Priority high → low:
OPENCLAW_LIVE_OPENAI_KEY        # Highest (runtime injection)
OPENAI_API_KEYS                 # Multi-key (comma-separated)
OPENAI_API_KEY                  # Standard single key
OPENAI_API_KEY_1                # Wildcard match OPENAI_API_KEY_*
OPENAI_API_KEY_2
Enter fullscreen mode Exit fullscreen mode

Smart rotation: Keys only rotate on 429 Too Many Requests or rate_limit errors. Other errors (auth failure, model not found) throw immediately — no wasted rotation attempts. This ensures:

  • High-concurrency pressure is automatically distributed
  • Non-rate-limit errors surface fast
  • Multiple teams sharing a key pool don't interfere

Cloud Models: OpenAI, Claude, Gemini, DeepSeek

OpenAI GPT-5 Series 

{
  agents: {
    defaults: {
      model: {
        primary: "openai/gpt-5.1-codex",
        fallbacks: [
          "openai/gpt-5.1-mini",
          "openai/gpt-5.1"
        ]
      },
      // models also serves as allowlist
      models: [
        "openai/gpt-5.1-codex",
        "openai/gpt-5.1",
        "openai/gpt-5.1-mini"
      ]
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

Anthropic Claude & Google Gemini

Set environment variables and specify in config:

# Anthropic Claude
export ANTHROPIC_API_KEY="sk-ant-xxxxx"

# Google AI Studio
export GOOGLE_API_KEY="AIzaSyxxxxx"

# Google Vertex AI (enterprise, needs service account)
export GOOGLE_VERTEX_PROJECT="my-project-id"
export GOOGLE_VERTEX_LOCATION="us-central1"
Enter fullscreen mode Exit fullscreen mode 
{
  agents: {
    defaults: {
      model: {
        primary: "anthropic/claude-opus-4-20250918",
        fallbacks: ["google/gemini-2.5-pro"]
      }
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

DeepSeek V3

DeepSeek uses OpenAI-compatible format via a custom Provider:

{
  models: {
    providers: {
      deepseek: {
        baseUrl: "https://api.deepseek.com/v1",
        apiKey: "sk-xxxxx"
      }
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

Then use deepseek/deepseek-chat or deepseek/deepseek-coder as model identifiers.

Local Models: Ollama & llama.cpp

For privacy-focused or offline scenarios, OpenClaw fully supports local models.

Ollama Setup

Ollama is the simplest local option:

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

# Pull models
ollama pull llama3.1:70b
ollama pull codellama:34b

# Start (default port 11434)
ollama serve
Enter fullscreen mode Exit fullscreen mode

OpenClaw config:

{
  models: {
    providers: {
      ollama: {
        baseUrl: "http://localhost:11434/v1",
        apiKey: "ollama"  // Ollama doesn't need a real key
      }
    }
  },
  memorySearch: {
    provider: "ollama"
  }
}
Enter fullscreen mode Exit fullscreen mode

Performance tip: 70B models need ~40GB VRAM (A100). 7B-13B models run smoothly on consumer GPUs (RTX 4090, 24GB).

llama.cpp Direct Connection

For lower-level control:

./llama-server -m models/codellama-34b-instruct.Q5_K_M.gguf \
  --host 0.0.0.0 --port 8080 \
  -c 8192 -ngl 99
Enter fullscreen mode Exit fullscreen mode 
{
  models: {
    providers: {
      "llama-cpp": {
        baseUrl: "http://localhost:8080/v1",
        apiKey: "none"
      }
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

Local vs. Cloud: Quick Comparison

Dimension Local Cloud
Latency Hardware-dependent, 50-200ms/token Network + inference, 20-80ms/token
Privacy Fully local, zero leak risk Data passes through third-party servers
Cost One-time hardware investment Per-token billing, higher long-term
Quality Limited by local compute (7B-70B) Access to strongest models
Best for Sensitive code, offline, high-frequency Complex reasoning, highest quality

Recommended: hybrid strategy. Local for daily coding (low cost), cloud for complex architecture decisions (best quality).

API Gateway: Unified Management

When you need multiple Providers, multiple keys, and automatic failover, an API gateway is the optimal solution.

Crazyrouter: 627+ Models with One Key

Crazyrouter is an API gateway built for AI developers. One key accesses OpenAI, Anthropic, Google, DeepSeek, Mistral, and 627+ more:

{
  models: {
    providers: {
      crazyrouter: {
        baseUrl: "https://crazyrouter.com/v1",
        apiKey: "sk-your-crazyrouter-key"
      }
    }
  },
  agents: {
    defaults: {
      model: {
        primary: "crazyrouter/claude-opus-4-20250918",
        fallbacks: [
          "crazyrouter/gpt-5.1-codex",
          "crazyrouter/gemini-2.5-pro"
        ]
      }
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

Or even simpler — override built-in Provider endpoints:

export OPENAI_API_KEY="sk-your-crazyrouter-key"
export OPENAI_BASE_URL="https://crazyrouter.com/v1"

export ANTHROPIC_API_KEY="sk-your-crazyrouter-key"
export ANTHROPIC_BASE_URL="https://crazyrouter.com/v1"
Enter fullscreen mode Exit fullscreen mode

Failover & Load Balancing

Combine fallbacks for automatic failover:

{
  agents: {
    defaults: {
      model: {
        primary: "crazyrouter/claude-opus-4-20250918",
        fallbacks: [
          "crazyrouter/gpt-5.1",
          "crazyrouter/gemini-2.5-pro",
          "crazyrouter/deepseek-chat"
        ]
      }
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

When the primary returns 429 or goes down, OpenClaw automatically tries the next model. Combined with key rotation, you get near 100% availability.

Three-layer fault tolerance architecture:

  1. Multiple keys per model auto-rotate (handles per-key rate limits)
  2. Multiple models auto-fallback (handles model-level failures)
  3. Gateway-level smart routing (handles provider-level outages)

Cost Optimization

API gateways also optimize costs:

  • Unified billing — no separate top-ups per Provider
  • Pay-as-you-go, no minimum spend
  • ~45% cost reduction vs. direct Provider connections
  • Built-in usage monitoring and budget alerts

For teams: shared key pool (no distributing individual Provider keys), per-project/member usage tracking, and daily/monthly budget caps.

Choosing Your Setup

Scenario Recommended
Solo dev, single Provider Direct env vars
Multi-model switching Fallbacks + multiple Providers
Privacy/offline Ollama local deployment
Team/enterprise API gateway for unified management

No matter which approach, OpenClaw's flexible architecture has you covered. For maximum models with minimum config, try Crazyrouter — one key, one endpoint, 627+ models, ready to go.

# 30-second setup
export OPENAI_BASE_URL="https://crazyrouter.com/v1"
export OPENAI_API_KEY="sk-your-crazyrouter-key"
openclaw models set crazyrouter/claude-opus-4-20250918
Enter fullscreen mode Exit fullscreen mode

👉 Sign up for Crazyrouter — free tier available

Top comments (0)