xAI Provider

The xAI provider connects GoClaw to Grok models via xAI’s gRPC API. It’s built on xai-go , a custom Go gRPC client for maximum performance.

Highlights

• 2M context window — Handle massive conversations, entire codebases, long documents
• Server-side tools — Web search, X search, code execution run on xAI infrastructure
• Stateful conversations — Context chaining reduces token usage and improves coherence
• Vision — Process images from Telegram, browser screenshots, etc.
• Reasoning — Configurable thinking levels for complex tasks
• Image generation — Create images via the xai_imagine tool
• gRPC streaming — Low-latency response streaming
• Competitive pricing — $0.20/$0.50 per million tokens (fast models)

Configuration

{
  "llm": {
    "providers": {
      "xai": {
        "driver": "xai",
        "apiKey": "YOUR_XAI_API_KEY"
      }
    },
    "agent": {
      "models": ["xai/grok-4-1-fast-reasoning"]
    }
  }
}

Options

Note: Context window sizes and pricing are fetched from the xAI API on startup. If the API is unreachable, hardcoded fallback values are used.

Models

Only Grok-4 family models are supported — they handle both vision and server-side tools simultaneously:

Pricing is per million tokens.

Note: Older models (grok-2, grok-3, grok-vision-beta) are blocked because they can’t handle both vision and server-side tools in the same request.

Server-Side Tools

xAI provides tools that execute on their infrastructure, not your machine:

All tools are enabled by default. To restrict:

{
  "providers": {
    "xai": {
      "driver": "xai",
      "apiKey": "YOUR_XAI_API_KEY",
      "serverToolsAllowed": ["web_search", "code_execution"]
    }
  }
}

How Server Tools Work

When the agent needs web search:

1. GoClaw sends the request to xAI
2. xAI’s infrastructure performs the search
3. Results stream back through the response
4. Agent sees the results and continues

This is faster than client-side tools and doesn’t require API keys for search services.

Tool Name Conflicts

When xAI server tools conflict with GoClaw client tools (both have web_search), the client tool is prefixed with local_:

• web_search — xAI server-side (fast, no API key needed)
• local_web_search — GoClaw client-side (uses your Brave API key)

The LLM sees both and chooses based on the task.

Stateful Conversations

xAI supports context preservation using previous_response_id. GoClaw manages this automatically:

Without incremental context:

Request 1: [system] + [msg1]
Request 2: [system] + [msg1] + [msg2]
Request 3: [system] + [msg1] + [msg2] + [msg3]  ← Tokens grow each turn

With incremental context (default):

Request 1: [system] + [msg1]              → responseID: abc123
Request 2: previousID: abc123 + [msg2]    → responseID: def456
Request 3: previousID: def456 + [msg3]    ← Only new messages sent

This dramatically reduces token usage for long conversations.

{
  "providers": {
    "xai": {
      "driver": "xai",
      "apiKey": "YOUR_XAI_API_KEY",
      "incrementalContext": true
    }
  }
}

Reasoning (Thinking)

Grok models support extended reasoning. Configure per-user:

{
  "users": [
    {
      "name": "Alice",
      "role": "owner",
      "thinking": true,
      "thinkingLevel": "medium"
    }
  ]
}

Higher levels take longer but produce better results for complex tasks.

Vision

Grok-4 models can process images. When you send an image via Telegram or the agent takes a browser screenshot, xAI sees and analyzes it.

No special configuration needed — just use a supported model.

Image Generation

GoClaw includes the xai_imagine tool for generating images:

{
  "tools": {
    "xaiImagine": {
      "enabled": true,
      "apiKey": "YOUR_XAI_API_KEY"
    }
  }
}

The agent can then generate images on request. See xAI Imagine Tool for details.

Agentic Turns

Limit how many turns the model can take when using server-side tools:

{
  "providers": {
    "xai": {
      "driver": "xai",
      "apiKey": "YOUR_XAI_API_KEY",
      "maxTurns": 5
    }
  }
}

This prevents runaway tool loops. Default is 25.

Troubleshooting

“Model not allowed”

Only Grok-4 family models work with GoClaw. Update your config:

"models": ["xai/grok-4-1-fast-reasoning"]

Server Tool Errors

Server-side tool failures are reported back to the LLM, which can retry or work around them.

Rate Limiting

The provider enters cooldown automatically. Check status:

/llm

Context Chain Broken

If the agent seems to “forget” context, the chain may have broken (e.g., after a compaction). It will rebuild automatically on the next request.

See Also

• xAI Imagine Tool — Image generation
• LLM Providers — Provider overview
• Configuration — Full config reference