Specification v0.7.0 · 37 Providers · V2 Protocol

The Specification
That Drives Everything.

AI-Protocol separates "what to do" from "how to do it." The V2 specification uses a three-layer pyramid with three-ring manifests. New in v0.7.0: MCP tool integration schema, Computer Use abstraction with safety policies, extended multimodal (audio/video/omni), ProviderContract, and a CLI validation tool.

Read the Docs View on GitHub

V2 Three-Layer Architecture

V2 organizes the specification into three layers: L1 Core Protocol (message format, error codes) → L2 Capability Extensions (streaming, tools, MCP, Computer Use, multimodal) → L3 Environment Profile (API keys, endpoints, retry). 6 V2 manifests available: OpenAI, Anthropic, Gemini, DeepSeek, Moonshot, Zhipu.

What's Inside

spec.yaml

Core Specification

Defines standard parameters (temperature, max_tokens), streaming events (PartialContentDelta, ToolCallStarted), error classes (13 types), and retry policies.

providers/

37 Provider Manifests

36 V1 + 6 V2 manifests. V2 manifests use three-ring structure with MCP, Computer Use, and multimodal capability declarations.

models/

Model Registry

Model instances with provider references, context windows, capability flags, and per-token pricing. GPT, Claude, Gemini, DeepSeek, Qwen, and more.

schemas/

V2 Schema Suite

6 JSON Schemas: provider, ProviderContract, MCP, Computer Use, multimodal, context policy. CI validation ensures correctness.

A Provider Manifest

Each provider is described by a YAML manifest. It declares everything a runtime needs to communicate with the provider — endpoint, authentication, parameter mapping, streaming decoder, error handling, and capabilities.

Runtimes read these manifests and "compile" user requests into provider-specific HTTP calls. No if provider == "openai" branches anywhere.

Endpoint & Auth — Base URL, protocol, bearer tokens, API key headers
Parameter Mapping — Standard names to provider-specific JSON fields
Streaming Decoder — SSE/NDJSON format, JSONPath event extraction rules
Error Classification — HTTP status codes to 13 standard error types

id: anthropic
protocol_version: "0.7.0"
endpoint:
  base_url: "https://api.anthropic.com/v1"
  chat_path: "/messages"
auth:
  type: bearer
  token_env: "ANTHROPIC_API_KEY"
parameter_mappings:
  temperature: "temperature"
  max_tokens: "max_tokens"
  stream: "stream"
streaming:
  decoder:
    format: "anthropic_sse"
  event_map:
    - match: "$.type == 'content_block_delta'"
      emit: "PartialContentDelta"
error_classification:
  by_http_status:
    "429": "rate_limited"
    "401": "authentication"
capabilities:
  streaming: true
  tools: true
  vision: true

Where Protocol Fits

AI-Protocol is the foundation layer. Runtimes consume it. Applications consume runtimes.

Supported Providers

Each provider has a complete YAML manifest with endpoint, auth, parameter mappings, streaming decoder, error handling, and capability flags.

OpenAI

Anthropic

Google Gemini

Groq

Mistral

DeepSeek

Qwen

Cohere

Azure OpenAI

Together AI

Perplexity

NVIDIA

Fireworks AI

Replicate

OpenRouter

DeepInfra

AI21 Labs

Cerebras

Lepton AI

Zhipu GLM

Doubao

Baidu ERNIE

Tencent Hunyuan

iFlytek Spark

Moonshot

MiniMax

Baichuan

Yi / 01.AI

SiliconFlow

SenseNova

Explore the Protocol

Read the specification, browse provider manifests, or contribute a new provider.

Documentation Contribute a Provider

The Specification That Drives Everything.