A Swift HTTP server that provides a REST interface to Apple's on-device Foundation Models (Apple Intelligence). Uses Hummingbird as the web framework.

• Swift 6.2 or newer
• macOS 26 or newer with Apple Intelligence available and enabled (compiler reports the Foundation Models APIs as macOS 26+)

swift build

Production mode:

swift run FoundationAPI

Debug mode:

swift run FoundationAPI --debug

GET /health

Returns ok if the server is running and the model is available.

POST /generate
Content-Type: application/json
{ "prompt": "your text" }

Returns JSON with both the input prompt and the model-generated response.

Example:

curl -X POST \
	-H "Content-Type: application/json" \
	-d '{"prompt":"Write a haiku about Xcode"}' \
	http://localhost:2929/generate

POST /v1/completions
Content-Type: application/json
{
 "prompt": "your text",
 "model": "auto",
 "max_tokens": 256,
 "temperature": 0.7,
 "stream": false
}

Returns OpenAI-compatible completion response format.

Example:

curl -X POST \
	-H "Content-Type: application/json" \
	-d '{
		"prompt": "Write a haiku about Xcode",
		"model": "auto",
		"max_tokens": 256,
		"temperature": 0.7
	}' \
	http://localhost:2929/v1/completions

The application supports runtime-configurable log levels:

• Normal mode (default): Only errors are displayed
• Debug mode: Includes detailed debug information about requests and model operations

Enable debug mode with the --debug flag when running:

swift run FoundationAPI --debug

Debug logging is controlled at runtime, so you don't need to rebuild to toggle debug output.

• Runs an HTTP server on port 2929 using Hummingbird.
• Creates a LanguageModelSession with concise instructions and temperature 0.7, maximum 256 response tokens.
• Returns JSON with the input prompt and the model response.
• Checks model availability and reports clear errors if the on-device model is unavailable, not enabled, or still downloading.