For AI agents: the complete documentation index is available at https://aioproxy.dev/llms.txt, the full documentation bundle is available at https://aioproxy.dev/llms-full.txt, and this page is available as Markdown at https://aioproxy.dev/guide/routing/protocols.md.
AIO Proxy
  • English

      • Protocols & Transparent Passthrough

      AIO Proxy supports diverse AI API protocols spanning chat completions, structured responses, vision, voice, embeddings, and agent runtimes.

      Supported Inbound Protocols

      Client ProtocolStandard Request PathFeature Highlights
      OpenAI Responses/v1/responsesNext-gen Responses API, streaming events, multimodal input, native tool lifecycle
      OpenAI Compatible/v1/chat/completionsStandard chat completions, function callings, tool calls, streaming SSE
      Anthropic Messages/v1/messagessystem prompt blocks, prompt caching, tool use, thinking tags
      Google Gemini/v1beta/models/*:generateContentGemini native contents format, multimodal inline data, safety settings
      OpenAI Embeddings/v1/embeddingsVector representations supporting batch text inputs
      OpenAI Realtime/v1/realtime, /v1/realtime/callsLow-latency WebRTC and WebSocket bidirectional audio/text streams
      OpenAI Media/v1/images/*, /v1/audio/*Image generation/edits, speech synthesis, and audio transcriptions

      Zero-Overhead Passthrough vs. Protocol Transforms

      AIO Proxy features a dual-engine architecture:

      1. Same-Protocol Raw Passthrough

      • When a client request matches the upstream provider's native protocol (e.g. Anthropic Client -> Anthropic Provider), AIO Proxy engages Raw Passthrough Mode.
      • Request bodies and response streams are piped transparently with zero-copy stream processing, introducing negligible latency (<1ms) while preserving vendor-specific headers, beta flags, and custom parameters.

      2. Cross-Protocol Intelligent Transforms

      • When protocols differ (e.g. Claude Code Anthropic Messages -> OpenAI Provider), AIO Proxy converts inbound payloads into universal model messages and invokes upstream via standardized adapter transforms.
      • Tools, function calling parameters, thoughts, and stream fragments are accurately translated.

      Multi-Protocol Endpoints

      Certain providers (such as OpenRouter, aggregator gateways, or multi-modal services) expose multiple protocol endpoints under one Base URL. You can declare endpoints to enable matching raw passthrough across different paths:

      config.jsonc
      {
  "providers": {
    "custom-gateway": {
      "kind": "api",
      "apiKey": "{{env.GATEWAY_KEY}}",
      "models": ["gpt-5", "claude-sonnet-4-6"],
      "endpoints": [
        // Serves OpenAI Compatible requests
        {
          "baseURL": "https://api.gateway.example/v1",
          "protocol": "openai-compatible",
        },
        // Serves Anthropic Messages requests natively
        {
          "baseURL": "https://api.gateway.example/v1",
          "protocol": "anthropic",
          "auth": "bearer",
        },
      ],
    },
  },
}

      With this configuration:

      • Inbound /v1/chat/completions requests will passthrough to the OpenAI-compatible endpoint.
      • Inbound /v1/messages requests will passthrough directly to the Anthropic endpoint.