Lite Agent
  • English

      • Supported protocols

      This page distinguishes model APIs, tool-call formats and MCP connections. See supported providers to choose a vendor. Every integration uses the same SDK constructor, createLiteAgent().

      Model service APIs

      These paths are available through @lite-agent/provider 0.11.0 and require Node.js >=22. AI SDK adapters must expose the AI SDK 7 LanguageModelV4 contract.

      API / protocol familyIntegrationTypical vendors / platformsStreaming transport
      OpenAI Chat Completionsopenai(), localOpenAI(), or aiSdk() with a Chat adapterOpenAI, DeepSeek, Qwen, GLM, Kimi and compatible endpointsHTTP + SSE
      OpenAI ResponsesaiSdk(createOpenAI()(modelId)); Azure uses its adapterOpenAI, Azure OpenAIHTTP + SSE
      xAI ResponsesaiSdk(createXai()(modelId))xAI GrokHTTP + SSE
      Anthropic Messagesanthropic() or aiSdk() with the corresponding adapterAnthropic; MiniMax uses Messages compatibilityHTTP + SSE
      Gemini / Vertex GeminiaiSdk() with Google or Vertex adaptersGoogle Gemini, Google VertexHTTP + SSE
      Amazon Bedrock ConverseaiSdk() with the Bedrock adapterBedrock models supporting ConverseHTTP + AWS binary event stream
      Cohere Chat v2aiSdk() with the Cohere adapterCohereHTTP + SSE

      Chat Completions and Responses are distinct APIs. The direct openai() adapter calls Chat Completions; choose an AI SDK Responses model for Responses. Authentication, fields and model capabilities can differ even within a protocol family; vendor adapters handle those differences.

      Perplexity uses its dedicated adapter and is currently covered for text. See the provider matrix for other vendors and packages. Protocol compatibility is not certification of every model, tool capability or parameter on a service.

      Tool-call formats (codecs)

      A codec describes tools to the model and decodes generated calls. This is independent of the HTTP API above.

      FormatConfigurationRequirements
      Native tool callingDefault nativeCodec()Model/API supports function calling; arguments are schema-validated before permission and hook processing
      JSON prompt formatcodec: jsonCodec()Model follows the JSON convention; tool definitions are embedded in the prompt
      ReAct prompt formatcodec: reactCodec()Model follows the ReAct format; selection is explicit

      Changing codecs does not automatically add model capabilities. Validate prompt-format reliability with the selected model. See Core tool-call codecs.

      MCP external tools

      MCP connects external tool servers; it is not a model-vendor API.

      ItemSupport
      Protocol versionPinned 2026-07-28, without automatic fallback
      Transportsstdio, Streamable HTTP
      AuthenticationHTTP headers; host-configured OAuth code + PKCE, refresh and machine credentials
      Agent integrationGlobal/project mcps.json, constructor mcpServers, agent.mcp

      Legacy HTTP+SSE transport and older MCP revisions are unsupported. See MCP servers for configuration and security boundaries.

      Capability and verification boundaries

      The AI SDK bridge supports multi-turn text, reasoning and function tools. It does not expose image/audio/video generation or enable provider-executed tools. LanguageModelV4 is an adapter interface version, not an HTTP protocol.

      Protocol tests, individual model capabilities and remote account access are separate checks. Explicit vendor reports of unsupported settings fail rather than silently ignoring tool or reasoning controls. See AI SDK integration and direct adapter settings.