This project provides a proxy server that allows Raycast AI to utilize models from any OpenAI-compatible API (OpenAI, Gemini, OpenRouter, etc.). This brings "Bring Your Own Key" (BYOK) functionality to Raycast AI, meaning you can use your own API key and models from your chosen provider. By default, the proxy is configured to use OpenRouter.
This proxy allows using custom models inside Raycast, including **AI Chat**, **AI Commands**, **Quick AI**, and **AI Presets**, giving you Raycast's native AI experience with the flexibility of custom models and your own API key.
## Features
This proxy aims to provide a seamless experience for using custom models within Raycast. Here's what is supported and what is not:
### Supported:
- 🧠 **Any model**: Access the wide range of models offered by OpenAI-compatible providers. OpenRouter is used by default.
- 👀 **Vision support**: Use models capable of processing images.
- 🛠️ **AI Extensions & MCP**: Use your favorite AI Extensions and MCP servers.
- 📝 **System instructions**: Provide system-level prompts to guide model behavior.
- 📎 **Attachments**: Attach all the same things as with the official models.
- 🔨 **Parallel tool use**: Make multiple tool calls simultaneously.
- ⚡ **Streaming**: Get real-time responses from models.
- 🔤 **Chat title generation**: Automatically generate chat titles.
- 🛑 **Stream cancellation**: Stop ongoing responses from models.
### Partial Support:
- 💭 **Displaying thinking process**: See the model's thinking process.
- This feature isn't supported by all providers because the OpenAI API specification does not define a standard for it. For example, when using OpenRouter, the thinking process is always shown by default for supported models. Other providers may not send it by default and require extra setup via the `extra` field in the model's configuration as described in the provider's documentation.
### Not Supported:
- 🌐 **Remote tools**: Some AI Extensions are classified as "remote tools" and are not supported. These include web search and image generation, as well as some others. You can replace these with MCP servers if you would like similar tools.
## Requirements
- Docker
- API key for your chosen provider (e.g., OpenRouter)
## Model Configuration
The proxy's behavior is primarily configured through a `models.json` file in the root directory of the project. This file defines the models available to Raycast and their specific settings. Each entry in the JSON array represents a model and can include the following properties:
-`name`: The name of the model as it will appear in Raycast.
-`id`: The model ID in the format expected by your provider.
-`contextLength`: The maximum context length (in tokens) the model supports. Only affects Raycast's UI and not the model itself.
-`capabilities`: An array of strings indicating the model's capabilities.
-`"vision"`: The model can process images.
-`"tools"`: The model supports AI Extensions and MCP (tool calling).
-`temperature`: (Optional) Controls the creativity of the model. A value between 0 and 2.
-`topP`: (Optional) Another parameter to control the randomness of the output, a value between 0 and 1.
-`max_tokens`: (Optional) The maximum number of tokens the model is allowed to generate in a single response.
-`extra`: (Optional) An object for advanced, provider-specific configurations. These options are passed directly to the provider's API.
### Default Configuration
The default models.json has 6 OpenRouter models configured, with the first four being free models and the last two being paid models. You can modify the models.json configuration according to your needs. Official OpenRouter model query:
