AI SDK

General definition

An AI SDK is a set of libraries, APIs and reference code that lets a development team add language-model features to a product without wiring up every model call, prompt, retrieval step and tool integration by hand. Where a bare LLM API returns text for a prompt, an AI SDK packages the surrounding machinery: conversation state, knowledge retrieval, tool execution, streaming, moderation and the connection to the place where users actually talk to the model.

• Agents: a model given a role, memory and a loop that decides when to answer, ask, or call a tool (see AI agent)
• Retrieval: indexing documents or websites so answers are grounded in your own content (RAG)
• Tool calling: letting the model invoke functions, APIs or MCP servers to look things up or take actions
• Streaming: sending tokens to the client as they are generated so replies feel immediate
• Model abstraction: swapping between cloud model providers and self-hosted models without rewriting application code

The term is used for two rather different kinds of product. Client-side libraries such as the Vercel AI SDK give a JavaScript application a uniform way to call model providers and render streamed output; they are model-access toolkits. Platform AI SDKs go further and supply the server-side pieces as well: agent runtime, knowledge base, memory, moderation and delivery into chat, voice or a web widget. Which one a team needs depends on whether the model call is the hard part or whether the product around it is.

For regulated buyers the deciding questions are usually where the model runs, where the retrieved data lives and what is logged. An AI SDK that supports self-hosted or private models and keeps transcripts on infrastructure the customer controls fits healthcare, finance and insurance use cases that a pure cloud-API toolkit cannot.

In the Ethora ecosystem

The Ethora AI SDK is the platform kind. Agents attach to chat rooms and channels, answer with RAG over a knowledge base built from a website crawler, uploaded files and per-agent live web indexing, call tools, and can be set to speak first on a schedule or after idle time. Each agent has its own system prompt and its own model choice: a cloud model API or a self-hosted model reached through an OpenAI-compatible endpoint such as Ollama or vLLM, which is the self-hosted LLM agent path for teams whose data cannot leave their environment.

Because the AI SDK shares one backend with the Chat SDK, an agent’s replies arrive as ordinary messages with threads, reactions and read receipts, and the same Trust & Safety rules, audit trail and retention settings apply. The MCP server lets Claude and other MCP clients create apps, users and rooms, and the embeddable assistant widget puts an agent on any website with a script tag.