Skip to main content

​
Why Choose Anthropic Native Mode

OpenClaw supports two ways to call Claude models. If you need tool calling (tool_use) and other advanced features, Anthropic native mode (anthropic-messages) is strongly recommended:
ScenarioOpenAI Compatible ModeAnthropic Native Mode
Basic Chat✅ Works✅ Works
Tool Calling (tool loop)❌ May return 400✅ Stable
Prompt Caching❌ Not supported✅ Supported
Multi-model Switching✅ 200+ models⚠️ Claude series only
With openai-completions, basic chat works fine, but multi-turn tool calling (tool_calls → tool_result → tool loop) may be rejected with a 400 error. Switching to anthropic-messages allows tool_use + tool_result format to work properly.
Edit ~/.openclaw/openclaw.json and add the following provider configuration: 
{
  "models": {
    "providers": {
      "apiyi": {
        "baseUrl": "https://api.apiyi.com",
        "apiKey": "sk-your-apiyi-key",
        "api": "anthropic-messages",
        "headers": {
          "anthropic-version": "2023-06-01",
          "anthropic-beta": ""
        },
        "models": [
          {
            "id": "claude-sonnet-4-6",
            "name": "Claude Sonnet 4",
            "reasoning": false,
            "contextWindow": 200000,
            "maxTokens": 16384
          },
          {
            "id": "claude-opus-4-6",
            "name": "Claude Opus 4",
            "reasoning": false,
            "contextWindow": 200000,
            "maxTokens": 16384
          }
        ]
      }
    }
  }
}

​
Critical Configuration Notes

The following three points must be set correctly, or you will encounter 400 errors:
  1. baseUrl without /v1: Must be https://api.apiyi.com — adding /v1 would result in .../v1/v1/messages causing request failure
  2. headers must include anthropic-version: Set to 2023-06-01
  3. anthropic-beta set to empty string: Disables beta feature headers to avoid triggering unsupported features

​
About reasoning: false

Claude models on APIYI will return a 400 error if the request contains thinking-related fields (thinking / output_config).Setting "reasoning": false in the model entry prevents OpenClaw from sending thinking fields, avoiding this issue.

​
Model Allowlist Configuration

Add models to agents.defaults.models, otherwise OpenClaw may report the model as “unregistered” and silently fall back to another model: 
{
  "agents": {
    "defaults": {
      "model": { "primary": "apiyi/claude-sonnet-4-6" },
      "models": {
        "apiyi/claude-sonnet-4-6": { "streaming": false },
        "apiyi/claude-opus-4-6": { "streaming": false }
      }
    }
  }
}

​
Comparison with OpenAI Compatible Mode

FeatureOpenAI Compatible ModeAnthropic Native Mode
API Typeopenai-completionsanthropic-messages
baseUrlhttps://api.apiyi.com/v1https://api.apiyi.com
Supported ModelsAll 200+ modelsClaude series only
Tool CallingUnstable (multi-turn may 400)Stable
Prompt CachingNot supportedSupported
Extended ContextDepends on modelUp to 200K tokens
Best ForMulti-model switching, basic chatDeep Claude usage, Agent tool calling

​
Claude Model ID List

Model IDNameDescription
claude-sonnet-4-6Claude Sonnet 4Balanced performance, recommended for daily use
claude-opus-4-6Claude Opus 4Strongest reasoning capability
claude-haiku-4-5-20251001Claude Haiku 4.5Fast response, cost-effective
Configure both OpenAI compatible and Anthropic native providers, switching as needed: 
{
  "agents": {
    "defaults": {
      "model": { "primary": "apiyi-claude/claude-sonnet-4-6" },
      "models": {
        "apiyi-claude/claude-sonnet-4-6": { "streaming": false },
        "apiyi-claude/claude-opus-4-6": { "streaming": false }
      }
    }
  },
  "models": {
    "providers": {
      "apiyi": {
        "baseUrl": "https://api.apiyi.com/v1",
        "apiKey": "sk-your-apiyi-key",
        "api": "openai-completions",
        "models": [
          { "id": "gpt-5.4", "name": "GPT-5.4" },
          { "id": "deepseek-v3.2", "name": "DeepSeek V3.2" }
        ]
      },
      "apiyi-claude": {
        "baseUrl": "https://api.apiyi.com",
        "apiKey": "sk-your-apiyi-key",
        "api": "anthropic-messages",
        "headers": {
          "anthropic-version": "2023-06-01",
          "anthropic-beta": ""
        },
        "models": [
          {
            "id": "claude-sonnet-4-6",
            "name": "Claude Sonnet 4",
            "reasoning": false,
            "contextWindow": 200000,
            "maxTokens": 16384
          },
          {
            "id": "claude-opus-4-6",
            "name": "Claude Opus 4",
            "reasoning": false,
            "contextWindow": 200000,
            "maxTokens": 16384
          }
        ]
      }
    }
  }
}
Use /model apiyi/gpt-5.4 or /model apiyi-claude/claude-sonnet-4-6 in chat to switch models.

​
Verify Configuration

After setup, verify that the configuration is working: 
{/* Check model status */}
openclaw models status

{/* Send a test message and check the returned provider/model */}
openclaw agent --message "just reply pong" --json
In the returned JSON, check that meta.agentMeta.provider and meta.agentMeta.model match your configuration.

​
Troubleshooting

This is usually caused by thinking-related fields in the request. Ensure:
  • Model entries have "reasoning": false set
  • Headers include "anthropic-beta": "" correctly configured
Existing chat sessions may have cached the old model configuration. Two solutions:Patch the session model:
openclaw gateway call sessions.patch \
  --params '{"key":"your-session-key","model":"apiyi-claude/claude-sonnet-4-6"}'
Or reset the session:
openclaw gateway call sessions.reset \
  --params '{"key":"your-session-key","reason":"reset"}'
Check if the model has been added to the agents.defaults.models allowlist. Unregistered models will be automatically fallen back by OpenClaw.
The Anthropic native mode baseUrl must not include /v1. Using https://api.apiyi.com/v1 will result in .../v1/v1/messages, causing a 404 error.