OpenCode

Run OpenCode against staik via the Anthropic endpoint — Swedish GPU infrastructure with prompt caching on reused context.

Configuration

Add staik as an Anthropic provider in ~/.config/opencode/opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "staik": {
      "npm": "@ai-sdk/anthropic",
      "name": "staik",
      "options": {
        "baseURL": "https://api.staik.se/v1",
        "apiKey": "{env:STAIK_API_KEY}"
      },
      "models": {
        "qwen3.6:35b-a3b": {
          "name": "Qwen 35B (staik)",
          "limit": { "context": 262144, "output": 32768 }
        },
        "qwen3.5:9b": {
          "name": "Qwen 9B (staik)",
          "limit": { "context": 262144, "output": 8192 }
        },
        "gemma4:31b": {
          "name": "Gemma 4 31B (staik)",
          "limit": { "context": 262144, "output": 32768 }
        }
      }
    }
  }
}

export STAIK_API_KEY=sk-st-your-key

The key is sent as x-api-key. baseURL points at /v1 — OpenCode appends /messages itself. No other changes needed.

Prompt caching

Agent loops resend the same large context (system prompt, repo files, history) every turn. Via the Anthropic endpoint, OpenCode marks the stable part with cache_control, and staik bills it as discounted cache_read instead of full price each time. You can see it in the response:

"usage": {
  "input_tokens": 101,
  "output_tokens": 2,
  "cache_creation_input_tokens": 0,
  "cache_read_input_tokens": 560
}

Details on TTL and breakpoints are in the Messages reference.

Models

qwen3.6:35b-a3b (default) has a 262K context window — enough for large repo contexts. qwen3.5:9b for fast simpler tasks, gemma4:31b for review. See Models.

Want to run Claude Code or the Anthropic SDK instead? See Claude Code.