# Zed AI setup

## Codex in the Agent Panel

The live configuration now has `disable_ai: false` and already registers `codex-acp`. Restart or reload Zed, open the ACP Registry, install or confirm Codex, then create a Codex thread from the Agent Panel. Codex owns its authentication and billing; use its native login flow for ChatGPT or API-based authentication supported by the installed version.

This does not require a VS Code extension. Zed hosts Codex through ACP.

## Self-hosted or compatible LLM API

Use an OpenAI-compatible provider for an endpoint implementing OpenAI chat-completions or Responses semantics:

```jsonc
"language_models": {
  "openai_compatible": {
    "my-provider": {
      "api_url": "https://llm.example.com/v1",
      "available_models": [
        {
          "name": "my-model",
          "display_name": "My Model",
          "max_tokens": 128000
        }
      ]
    }
  }
}
```

Enter the key through **Agent Settings > LLM Providers**, or set `MY_PROVIDER_API_KEY`. Do not put the key in `settings.json`.

Anthropic-compatible endpoints can be configured under `language_models.anthropic_compatible`. First-class providers include Anthropic, OpenAI, Google AI, Mistral, DeepSeek, xAI, and OpenCode. Zed also supports OpenRouter, Vercel AI Gateway, Bedrock, Ollama, LM Studio, llama.cpp, and local OpenAI-compatible servers.

Using your own key or local endpoint does not require paying for Zed-hosted models. The API provider may still charge for usage. Zed-hosted models use Zed's plan and usage billing.