Zed

Settings#

In the Agent Panel's settings, Add Provider in the LLM Providers section asks for a provider name, the API URL, a model ID and the context window. The same in settings.json, with more control:

{
  "language_models": {
    "openai_compatible": {
      "aitokens": {
        "api_url": "https://api.aitokens.ch/v1",
        "available_models": [
          {
            "name": "qwen3.8-27b",
            "display_name": "Qwen3.8 27B",
            "max_tokens": 262144,
            "max_output_tokens": 8192,
            "capabilities": {
              "tools": true,
              "images": false,
              "parallel_tool_calls": false,
              "prompt_cache_key": false,
              "chat_completions": true
            }
          },
          {
            "name": "qwen3.6-35b",
            "display_name": "Qwen3.6 35B",
            "max_tokens": 32768,
            "max_output_tokens": 8192,
            "capabilities": {
              "tools": true,
              "images": true,
              "parallel_tool_calls": false,
              "prompt_cache_key": false,
              "chat_completions": true
            }
          }
        ]
      }
    }
  }
}

In Zed, max_tokens is the context window and max_output_tokens the most the model writes in one answer.

The key, sk-at-…, does not go in settings.json. Paste it in the provider's settings, or set the environment variable Zed derives from the provider ID: the ID in capitals with _API_KEY after it.

Capabilities#

• chat_completions: true: keep it. AI Tokens's Responses API has no streaming, and Zed would use it for a model with false.
• prompt_cache_key: false: keep it. The API does not know that field and refuses a request that carries it with 400.
• images: true only on models that read images, such as qwen3.6-35b.

Reasoning#

qwen3.8-27b reasons before every answer, and Zed has nothing to show of it: the API does not return it. To switch it off, add "reasoning_effort": "none" to the model in settings.json. The API accepts none, minimal, low, medium, high and xhigh; Zed's max is refused.

Limits#

• No edit predictions. Zed's edit predictions through an OpenAI-compatible API use fill-in-the-middle on /v1/completions, which AI Tokens does not serve. The agent uses chat completions and is not affected.
• max_output_tokens up to 8192. The API refuses a higher limit with 422.