reference.md

aimux API reference

Lookup index for the public API surface. Rust is the source of truth; TypeScript declarations in bindings/node/src/types/ are generated by ts-rs and mirror the JSON shapes that flow through the FFI boundary (all bindings use the same shapes). Full provider list: providers.md.

Types

Type	Rust source	TS declaration	Notes
GenerateTextOptions	generate.rs	GenerateTextOptions.ts	per-call text options (temperature, tools, reasoning, timeout, …)
CallOptions	options.rs	CallOptions.ts	base options for all modalities
TimeoutConfiguration	options.rs	TimeoutConfiguration.ts	total_ms / step_ms / first_chunk_ms / chunk_ms
ResponseFormat	options.rs	ResponseFormat.ts	text / json_object / json_schema
ReasoningEffort	types.rs	ReasoningEffort.ts	7 levels, passed through verbatim
AbortSignal	abort_signal.rs	— (runtime handle)	Rust: options field; Node: AbortBridge + JS AbortSignal
aimux_providers::provider::ProviderOptions (binding compatibility)	provider.rs	ProviderConfig (per binding)	base_url / headers / organization / project / params (a max_retries or body_overrides key is rejected as InvalidArgument)
Output	output.rs	Output.ts	the output option: Text / Object / Array / Choice / Json
GenerateTextResult	generate.rs	GenerateTextResult.ts	steps, text, content, tool_calls, usage, finish_reason, output, …
StepResult	step_result.rs	StepResult.ts	one model call: content, usage, performance, request, response
StreamTextResult	generate.rs	—	use TextStreamPart while iterating; consume() gives a GenerateTextResult
TextStreamPart	stream_part.rs	TextStreamPart.ts	Start / StartStep / TextDelta / ToolCall / FinishStep / Finish / …
StreamPart	stream_part.rs	StreamPart.ts	the provider-level stream (LanguageModel::do_stream)
LanguageModelUsage	types.rs	LanguageModelUsage.ts	call-level usage: flat input/output/total tokens plus details
Usage / TokenUsage	types.rs	Usage.ts	provider-level usage
FinishReason / FinishReasonUnified	types.rs	FinishReason.ts
Warning	types.rs	Warning.ts	non-fatal provider notices
AiMuxError	error.rs	AiMuxError.ts	typed error with error_type + status_code
ModelMessage / Role / MessageContent / ModelPrompt	message.rs	ModelMessage.ts	chat message shapes
LanguageModelPromptMessage	language_model_message.rs	LanguageModelPromptMessage.ts	canonical prompt message
ContentPart	content.rs	ContentPart.ts	Text / Reasoning / ToolResult / …
Tool / FunctionTool / ProviderTool	tool.rs	Tool.ts	tool definitions
ToolCall / ToolResult / ToolChoice	tool.rs	ToolCall.ts	tool invocation plumbing
Embedding types	embedding_model.rs	Embedding*.ts	EmbeddingCallOptions / EmbeddingResult / EmbeddingResponse
Image types	image_model.rs	Image*.ts	ImageCallOptions / ImageResult / ImageResponse / ImageFile
Speech types	speech_model.rs	Speech*.ts	SpeechCallOptions / SpeechResult / AudioData
Transcription types	transcription_model.rs	Transcription*.ts	TranscriptionCallOptions / TranscriptionResult / TranscriptionSegment
Video types	video_model.rs	Video*.ts	VideoCallOptions / VideoResult / VideoData
Reranking types	reranking_model.rs	Reranking*.ts	RerankingCallOptions / RerankingResult / RerankingRank
Search types	search_model.rs	Search*.ts	SearchCallOptions / SearchResult / SearchResultItem
Files types	files_model.rs	UploadFile*.ts	UploadFileCallOptions / UploadFileResult
Shared	shared.rs	FileBytes.ts / FileData.ts / Size.ts / AspectRatio.ts

Functions

Native protocols

Provider	Rust	Node / Python	Go / Java	Kotlin / Swift / Flutter	C ABI
OpenAI	create_openai(OpenAIProviderSettings { api_key: Some(key.into()), ..Default::default() })	openai(apiKey, model, baseUrl?)	NewOpenAI(...) / Model.openai(...)	Model.openai(...) / Aimux.openai(apiKey:modelId:)	aimux_openai_new
Anthropic	create_anthropic(AnthropicProviderSettings { api_key: Some(key.into()), ..Default::default() })	anthropic(apiKey, model, baseUrl?)	NewAnthropic(...) / Model.anthropic(...)	Model.anthropic(...) / Aimux.anthropic(apiKey:modelId:)	aimux_anthropic_new
Google	GoogleProvider	google(apiKey, model, baseUrl?)	—	—	—
DeepSeek	aimux_providers::deepseek::create_deepseek(DeepSeekProviderSettings::default())	deepseek(...)	NewDeepSeek(...) / Model.deepseek(...)	—	aimux_provider_new("deepseek", ...)
Cohere	CohereProvider	cohere(apiKey, model, baseUrl?)	NewCohere(...) / Model.cohere(...)	Model.cohere(...) / Aimux.cohere(...)	aimux_cohere_new
Mistral	MistralProvider	mistral(apiKey, model, baseUrl?)	NewMistral(...) / Model.mistral(...)	Model.mistral(...) / Aimux.mistral(...)	aimux_mistral_new
xAI	XAIProvider	xai(apiKey, model, baseUrl?)	NewXai(...) / Model.xai(...)	Model.xai(...) / Aimux.xai(...)	aimux_xai_new
Bedrock	AmazonBedrockProvider	bedrock(accessKeyId, secretKey, region, model, baseUrl?)	NewBedrock(...) / Model.bedrock(...)	Model.bedrock(...) / Aimux.bedrock(...)	aimux_bedrock_new
Vertex	VertexProvider	vertex(accessToken, project, location, model, baseUrl?)	NewVertex(...) / Model.vertex(...)	Model.vertex(...) / Aimux.vertex(...)	aimux_vertex_new
Anthropic-AWS	AnthropicAwsProvider	anthropicAws(apiKey, region, model, baseUrl?)	NewAnthropicAws(...) / Model.anthropicAws(...)	Model.anthropicAws(...) / Aimux.anthropicAws(...)	aimux_anthropic_aws_new
Azure	AzureOpenAIProvider	azure(apiKey, resource, deployment, apiVersion?, baseUrl?)	NewAzure(...) / Model.azure(...)	Model.azure(...) / Aimux.azure(...)	aimux_azure_new

— = not exposed in that binding yet (Google is Rust / Node / Python only). Every constructor has a base-URL variant (WithBase / baseUrl overload).

Other vendor packages (speech, transcription, image, video and search) have typed Rust factories; presets use create_provider(name, PresetSettings): providers.md. Their per-binding modality factories are in the function tables below.

Core (Rust, aimux_providers / aimux_core)

Function	Signature (abridged)	Notes
create_provider	(name, PresetSettings) -> Result<Arc<dyn Provider>, AiMuxError>	vendor package or preset; unknown name → NoSuchProvider; keys read per request
default_providers	() -> BTreeMap<String, Arc<dyn Provider>>	all built-in providers with default settings
provider_names	() -> Vec<&'static str>	sorted names accepted by create_provider
provider_registry::create_provider_registry	(providers, options) -> ProviderRegistry	resolves "provider:model" ids; options set the separator
Provider::discovery	() -> Option<&dyn ProviderDiscovery>	optional model listing via list_models().await
generate_text	(model, prompt, options) -> GenerateTextResult	non-streaming
stream_text	(model, prompt, options) -> StreamTextResult	streaming

Per-modality entry points are trait methods on the model types (EmbeddingModel, ImageModel, SpeechModel, …) — see API.md §Features.

The bindings keep their own compatibility API, implemented in aimux_providers::provider; these helpers and option types are not Rust crate-root exports.

Node (@arcships/aimux)

Function	Notes
generateText(model, prompt, options?, signal?) / streamText(...)	typed wrappers; signal = AbortSignal
openai() / anthropic() / deepseek() / google()	native typed path; deepseek() uses the vendor chat model
provider(name, apiKey?, modelId, config?)	by-name vendor packages and presets; names listed in providers.md
openaiEmbedding / cohereEmbedding / googleEmbedding	embeddings
openaiSpeech / openaiTranscription	speech / STT
openaiImage / googleImage / googleVideo	image / video
openaiFiles / cohereReranking / tavilySearch	files / rerank / search
AbortBridge	wraps a JS AbortSignal for the raw napi API

Python (aimux)

Function	Notes
generate_text(model, prompt, options?) / stream_text(...)	typed wrappers
openai / anthropic / deepseek / provider	same semantics as Node; deepseek uses the vendor chat model
openai_embedding / cohere_embedding / google_embedding	embeddings
openai_speech / openai_transcription	speech / STT
openai_image / google_image / google_video	image / video
openai_files / cohere_reranking / tavily_search	files / rerank / search

Go (github.com/arcships/aimux/bindings/go)

Function	Notes
Provider(name, apiKey?, modelId, config?)	by-name vendor packages and presets; names listed in providers.md
NewOpenAI / NewAnthropic / NewDeepSeek	native typed path (+ WithBase variants)
NewOpenAIEmbedding / NewCohereEmbedding / NewGoogleEmbedding	embeddings
NewOpenAISpeech / NewOpenAITranscription	speech / STT
NewOpenAIImage / NewGoogleImage / NewGoogleVideo	image / video
NewOpenAIFiles / NewCohereReranking / NewTavilySearch	files / rerank / search

Java / Kotlin / Flutter (Model.*) and Swift (Aimux.*)

Function	Notes
provider(name, apiKey?, modelId, config?)	by-name vendor packages and presets
openai(apiKey, modelId) / anthropic(apiKey, modelId)	native typed path (+ WithBase/baseUrl variants; Swift: Aimux.openai(apiKey:modelId:))
modality classes	EmbeddingModel / SpeechModel / ImageModel / TranscriptionModel / VideoModel / RerankingModel / SearchModel / Files, each with openai / cohere / google … factories

C ABI (aimux-ffi)

All functions communicate via JSON strings. Entry points are declared in aimux-ffi.h: provider_new (by-name vendor package or preset), model constructors (openai_new, anthropic_new, …), *_generate, *_stream (push-callback), *_embed / *_speech / … per modality, and the *_free family.

Providers

The complete list — registry-backed (name / display / env var / base URL) and typed-factory providers grouped by category, with the canonical counts — is in providers.md. It is generated from aimux-providers/src/provider_registry.json + aimux-providers/src/lib.rs: python scripts/gen_providers_doc.py.