Skip to content
AI-Lib

Introduction

Welcome to AI-Lib

Section titled “Welcome to AI-Lib”

AI-Lib is an open-source ecosystem that standardizes how applications interact with AI models. Instead of writing provider-specific code for each AI service, you use a single unified API — and protocol configuration handles the rest.

The Core Idea

Section titled “The Core Idea”

All logic is operators, all configuration is protocol.

Traditional AI SDKs embed provider-specific logic in code: different HTTP endpoints, different parameter names, different streaming formats, different error codes. When you switch providers, you rewrite code.

AI-Lib takes a different approach:

  • AI-Protocol defines how to talk to each provider in YAML manifests
  • Runtime implementations (Rust, Python) read these manifests and execute requests
  • Zero hardcoded logic — no if provider == "openai" branches anywhere

Three Projects, One Ecosystem

Section titled “Three Projects, One Ecosystem”
ProjectRoleLanguageVersionDistribution
AI-ProtocolSpecification layerYAML/JSONv0.7.0GitHub
ai-lib-rustRuntime implementationRustv0.8.0Crates.io
ai-lib-pythonRuntime implementationPythonv0.7.0PyPI

The latest release delivers the full V2 protocol with three major new capabilities: MCP tool integration, Computer Use abstraction, and Extended Multimodal support. Both runtimes feature ProviderDriver abstraction, Capability Registry, and 230+ tests ensuring cross-runtime consistency.

AI-Protocol (Specification)

Section titled “AI-Protocol (Specification)”

The foundation. YAML manifests describe 37 AI providers (6 V2 + 36 V1): their endpoints, authentication, parameter mappings, streaming decoder configurations, error classification rules, MCP/CU/multimodal capabilities, and more. JSON Schema validates everything.

ai-lib-rust (Rust Runtime)

Section titled “ai-lib-rust (Rust Runtime)”

High-performance runtime. Operator-based streaming pipeline processes responses through composable stages (Decoder → Selector → Accumulator → EventMapper). Built-in resilience with circuit breaker, rate limiter, and backpressure. Published on Crates.io.

ai-lib-python (Python Runtime)

Section titled “ai-lib-python (Python Runtime)”

Developer-friendly runtime. Full async/await support, Pydantic v2 type safety, production-grade telemetry (OpenTelemetry + Prometheus), and intelligent model routing. Published on PyPI.

Key Features

Section titled “Key Features”
  • 37 providers — OpenAI, Anthropic, Gemini, DeepSeek, Qwen, Moonshot, Zhipu, and many more (6 V2 + 36 V1)
  • Unified streaming — Same StreamingEvent types regardless of provider
  • Protocol-driven — All behavior defined in YAML, not code
  • MCP integration — Built-in MCP tool bridge: convert MCP server tools to AI-Protocol format automatically
  • Computer Use — Normalized GUI automation abstraction with safety policy enforcement
  • Extended multimodal — Vision, audio, video input; audio and image output; omni-mode support
  • ProviderDriver — Three concrete drivers (OpenAI, Anthropic, Gemini) with automatic API style detection
  • Capability Registry — Dynamic module loading based on manifest capability declarations
  • CLI tool — ai-protocol-cli for manifest validation, provider info, and compatibility checking
  • Hot-reload — Update provider configs without restarting
  • Resilience — Circuit breaker, rate limiting, retry, fallback
  • Tool calling — Unified function calling across providers
  • Embeddings — Vector operations and similarity search
  • Type safety — Compile-time (Rust) and runtime (Pydantic) validation

Next Steps

Section titled “Next Steps”