Download and install Ollama from [ollama.com/download](https://ollama.com/download) (Linux or macOS) and ensure it's running with `ollama --version`.

1. Download one of the [available models](https://ollama.com/models), for example, for `mistral`:

   ```sh
   ollama pull mistral
   ```

2. Make sure that the Ollama server is running. You can start it either via running Ollama.app (macOS) or launching:

   ```sh
   ollama serve
   ```

3. In the Agent Panel, select one of the Ollama models using the model dropdown.

#### Ollama Context Length {#ollama-context}

Zed has pre-configured maximum context lengths (`max_tokens`) to match the capabilities of common models.
Zed API requests to Ollama include this as `num_ctx` parameter, but the default values do not exceed `16384` so users with ~16GB of ram are able to use most models out of the box.
See [get_max_tokens in ollama.rs](https://github.com/zed-industries/zed/blob/main/crates/ollama/src/ollama.rs) for a complete set of defaults.

> **Note**: Token counts displayed in the Agent Panel are only estimates and will differ from the model's native tokenizer.

Depending on your hardware or use-case you may wish to limit or increase the context length for a specific model via settings.json:

```json
{
  "language_models": {
    "ollama": {
      "api_url": "http://localhost:11434",
      "available_models": [
        {
          "name": "qwen2.5-coder",
          "display_name": "qwen 2.5 coder 32K",
          "max_tokens": 32768,
          "supports_tools": true,
          "supports_thinking": true,
          "supports_images": true
        }
      ]
    }
  }
}
```

If you specify a context length that is too large for your hardware, Ollama will log an error.
You can watch these logs by running: `tail -f ~/.ollama/logs/ollama.log` (macOS) or `journalctl -u ollama -f` (Linux).
Depending on the memory available on your machine, you may need to adjust the context length to a smaller value.

You may also optionally specify a value for `keep_alive` for each available model.
This can be an integer (seconds) or alternatively a string duration like "5m", "10m", "1h", "1d", etc.
For example, `"keep_alive": "120s"` will allow the remote server to unload the model (freeing up GPU VRAM) after 120 seconds.

The `supports_tools` option controls whether or not the model will use additional tools.
If the model is tagged with `tools` in the Ollama catalog this option should be supplied, and built in profiles `Ask` and `Write` can be used.
If the model is not tagged with `tools` in the Ollama catalog, this option can still be supplied with value `true`; however be aware that only the `Minimal` built in profile will work.

The `supports_thinking` option controls whether or not the model will perform an explicit “thinking” (reasoning) pass before producing its final answer.  
If the model is tagged with `thinking` in the Ollama catalog, set this option and you can use it in zed.

The `supports_images` option enables the model’s vision capabilities, allowing it to process images included in the conversation context.  
If the model is tagged with `vision` in the Ollama catalog, set this option and you can use it in zed.

### OpenAI {#openai}

> ✅ Supports tool use

1. Visit the OpenAI platform and [create an API key](https://platform.openai.com/account/api-keys)
2. Make sure that your OpenAI account has credits
3. Open the settings view (`agent: open configuration`) and go to the OpenAI section
4. Enter your OpenAI API key

The OpenAI API key will be saved in your keychain.

Zed will also use the `OPENAI_API_KEY` environment variable if it's defined.

#### Custom Models {#openai-custom-models}

The Zed Assistant comes pre-configured to use the latest version for common models (GPT-3.5 Turbo, GPT-4, GPT-4 Turbo, GPT-4o, GPT-4o mini).
To use alternate models, perhaps a preview release or a dated model release, or if you wish to control the request parameters, you can do so by adding the following to your Zed `settings.json`: