llm-go

llm-go is a provider-neutral Go library for model generation, streaming, tool use, multimodal messages, reasoning, and usage accounting. It keeps provider SDK types behind a small common API so applications can switch protocols without changing their conversation model. The library requires Go 1.27 or newer, and its public API is still under active development.

It separates wire-protocol adapters from model and provider metadata. Every adapter supports generation, typed stream events, classified errors, and token usage. Built-in protocols include OpenAI Responses, Azure OpenAI Responses, OpenAI-compatible Chat Completions, ChatGPT subscription Codex Responses, Anthropic-compatible Messages, Gemini Developer API, Google Vertex AI, Amazon Bedrock ConverseStream, and Mistral Conversations. Model-scoped features must still be declared by the configured model.

Portable orchestration helpers are opt-in and do not change provider defaults. These include bounded retries, stream collection, structural collection, tool strictness, history transformation, token budgeting, tool argument validation, context-overflow detection, and JSON repair helpers. Tool inputs use JSON Schema 2020-12 by default and support declared drafts and local references offline.

Custom protocols and dynamic catalogs are supported through an immutable factory registry and a catalog manager that publishes a static baseline plus validated provider overlays. Built-in provider profiles supply protocol and endpoint defaults for services such as OpenRouter, Groq, DeepSeek, xAI, Cerebras, Fireworks, Mistral, Google Vertex, and Amazon Bedrock. Local servers exposing an OpenAI-compatible /v1/chat/completions endpoint use the openai adapter.