AI-Lib
🇺🇸 English 🇨🇳 简体中文 🇯🇵 日本語 🇪🇸 Español
Open Source Ecosystem

One Protocol,
Every AI Provider.

AI-Lib is an open-source ecosystem that standardizes AI model interactions. V2 protocol with MCP, Computer Use, and multimodal capabilities — 37 providers, three runtimes, zero hardcoded logic.

Get Started View on GitHub
37
AI Providers
5
Projects
0
Hardcoded Logic
3
Runtimes

Six Projects, One Ecosystem

A specification layer defines the rules. Four runtime implementations (Rust, Python, TypeScript, Go) bring them to life.

v1.0.0

AI-Protocol

The provider-agnostic specification. V2 three-ring manifests with MCP, Computer Use, and multimodal schemas — no hardcoded logic, ever.

  • 37+ provider manifests (10 V2 + 27 V1)
  • V2 three-layer pyramid architecture
  • STT / TTS / Rerank in manifests (Jina, OpenAI, Cohere)
  • MCP / Computer Use / Multimodal schemas
  • ProviderContract + pricing schema
  • CLI tool for validation & inspection
  • JSON Schema validated
  • Hot-reloadable configuration
Learn more
v1.0.1

ai-lib-rust

High-performance Rust runtime. E/P workspace (ai-lib-core + ai-lib-contact), manifest-driven Pipeline + AiClient, structured output, and feature-gated capability modules.

  • AiClient + operator-based streaming pipeline
  • ai-lib-core / ai-lib-contact workspace crates
  • 13 V2 standard error codes
  • Structured output & text-tool (TTC)
  • Optional: embeddings, MCP bridge, computer_use (feature-gated)
  • Opt-in resilience (ai_lib_rust::resilience)
  • Published on Crates.io v1.0.1
Learn more
v1.0.0

ai-lib-python

Async Python runtime. E/P module separation, manifest-driven Pipeline + AiClient, Pydantic v2 types, and pip-extra capability modules.

  • AiClient + operator-based streaming pipeline
  • Fluent chat builder (.messages / .user / .stream)
  • 13 V2 standard error codes
  • Structured output & text-tool types
  • Optional: embeddings, MCP bridge, computer_use (pip extras)
  • Opt-in resilience (production_ready / resilience module)
  • Published on PyPI v1.0.0
Learn more
v1.0.0

ai-lib-ts

TypeScript runtime for npm. HttpTransport + manifest parsers, /core and /contact entry points, opt-in policy layer.

  • AiClient + manifest-driven HTTP chat
  • /core (E) and /contact (P) entry points
  • 13 V2 standard error codes
  • StreamingEvent with event_type discriminator
  • Optional: EmbeddingClient, McpToolBridge (format only)
  • Default transport retry; CB/RL opt-in
  • Published on npm v1.0.0
Learn more
v1.0.0

ai-lib-go

Go runtime (v1.0.0). pkg/ailib execution client + pkg/contact fallback policy; protocol-first HTTP with ExecutionMetadata.

  • pkg/ailib Client + manifest HTTP chat
  • pkg/contact FallbackClient (circuit breaker)
  • 13 V2 standard error codes
  • ChatStream SSE decoder (openai_sse default)
  • Capability-gated endpoints (MCP/CU routes)
  • Go 1.21+ context-aware API
Learn more

Dev Tools

v1.0.1

ai-protocol-mock

Unified mock server for AI-Protocol runtimes. HTTP provider mock (OpenAI + Anthropic), STT/TTS/Rerank endpoints, and MCP JSON-RPC mock for integration testing.

  • Manifest-driven HTTP mock
  • OpenAI & Anthropic + STT/TTS/Rerank
  • MCP tools/list, tools/call, capabilities
  • Docker support
  • Third-party integration (ZeroClaw, CI)
  • Sync manifests from ai-protocol
View on GitHub →

Showcase Projects

Reference applications built on the AI-Lib ecosystem — see the protocol and runtimes in action.

v0.3.1

AI Debate

Multi-model AI debate arena. Pro vs Con across four rounds, then a Judge delivers the verdict. Built on ai-lib-rust and ai-protocol.

  • 4-round debate flow (Opening → Rebuttal → Defense → Closing → Judgement)
  • Web search tool calling via Tavily (optional)
  • Multi-provider: DeepSeek, Zhipu, Groq, Mistral, OpenAI, Anthropic
  • Auto fallback, real-time SSE streaming
  • Axum + SQLite, modern dark UI
View on GitHub →
v0.2.0

ZeroSpider

Protocol-driven autonomous AI agent runtime, forked from ZeroClaw with ai-protocol integration, smart routing, and multi-model negotiation.

  • Forked from ZeroClaw with ai-protocol ecosystem integration
  • Smart routing: cost, speed, quality, reliability scoring
  • Multi-model negotiation and parallel task execution
  • Channels: Telegram, Discord, Matrix
  • Remote deployment, hardware (GPIO, STM32)
View on GitHub →
v0.4.2

SpiderSwitch

MCP server for dynamic AI model switching. Enables agents to discover, switch, and manage AI models across providers via standard MCP tools.

  • MCP-compliant server over stdio transport
  • switch_model / list_models / get_status / exit_switcher tools
  • Protocol-driven: all configs from ai-protocol manifests
  • Multi-provider: OpenAI, Anthropic, Google, DeepSeek, etc.
  • Python + ai-lib-python SDK, auto protocol setup
View on GitHub →

Protocol-Driven Design

"All logic is operators, all configuration is protocol." Every provider behavior is declared in YAML — runtimes contain zero hardcoded provider logic.

Declarative Configuration

Provider endpoints, auth, parameter mappings, streaming decoders, and error handling — all declared in YAML manifests, validated by JSON Schema.

Operator-Based Pipeline

Streaming responses flow through composable operators: Decoder, Selector, Accumulator, FanOut, EventMapper. Each operator is protocol-driven.

Hot-Reload Ready

Update provider configurations without restarting. Protocol changes propagate automatically to runtimes. Add new providers through configuration, not code.

Ecosystem Architecture

Three layers working together — specification defines the rules, runtimes execute them, applications consume unified interfaces.

AI-Lib Ecosystem Architecture APPLICATION RUNTIME PROTOCOL Web Apps / API Services Rust / Python / TS Your application code AI Agents Multi-turn / Tool Calling CLI Tools Batch / Data Pipelines Testing & Dev Mock Services ai-lib-rust v1.0.1 AiClient Pipeline Transport Resilience Embeddings Cache / Batch Crates.io · tokio + reqwest · <1ms overhead ai-lib-python v1.0.0 AiClient Pipeline Transport Resilience Telemetry Routing PyPI · httpx + Pydantic v2 · async/await ai-lib-ts v1.0.0 AiClient Pipeline Transport Resilience MCP Bridge Batch NPM · fetch + Zod · ESM / CJS Load Manifests AI-Protocol v1.0.0 spec.yaml Core Specification providers/*.yaml 37 Provider Manifests models/*.yaml Model Registry YAML definitions → JSON compilation → Runtime consumption · Vendor neutral ai-protocol-mock v1.0.1 Mock Server Local Testing Record/Replay Snapshot Tests Protocol Compliance Check Dev Testing · CI/CD · No real API needed mock

How It Works

From user request to unified streaming events — every step is protocol-driven.

Request-Response Data Flow REQUEST FLOW → User chat() AiClient UnifiedRequest Protocol compile_request() Transport HTTP POST AI Provider OpenAI, etc. ← RESPONSE FLOW (SSE/JSON) Byte Stream SSE / NDJSON Decoder JSON Frames Pipeline Select → Accumulate → Map EventMapper StreamingEvent Application Unified Events All parameter mapping and event normalization driven by protocol manifests — zero hardcoded logic Unified events: PartialContentDelta · ToolCallStarted · PartialToolCall · StreamEnd

Choose Your Runtime

Same protocol, different strengths. Pick the runtime that fits your stack.

Capability AI-Protocol Rust SDK Python SDK TypeScript SDK Go SDK
Type System JSON Schema Compile-time (Rust) Runtime (Pydantic v2) Compile-time (TypeScript) Compile-time (Go Structs)
Streaming SSE/NDJSON spec tokio async streams async generators AsyncIterator + fetch Native stream iteration
Resilience Retry policy spec Circuit breaker, rate limiter, back-pressure ResilientExecutor with all patterns RetryPolicy, CircuitBreaker, RateLimiter Context timeouts, auto-retry
V2 Driver ProviderContract spec Box<dyn ProviderDriver> ProviderDriver ABC ManifestV2 + HttpTransport ProviderDriver interface
MCP mcp.json schema McpToolBridge McpToolBridge McpToolBridge To be implemented
Computer Use computer-use.json schema ComputerAction + SafetyPolicy ComputerAction + SafetyPolicy — —
Multimodal multimodal.json schema MultimodalCapabilities MultimodalCapabilities SttClient, TtsClient, RerankerClient MultimodalCapabilities
Embeddings — Vector operations, similarity Vector operations, similarity EmbeddingClient EmbeddingClient
Distribution GitHub / npm Crates.io PyPI npm goproxy
Best For Specification & standards Systems, performance-critical ML, data science, prototyping Node.js, npm ecosystem, full-stack High-concurrency servers, microservices

37 AI Providers Supported

All driven by protocol configuration — zero hardcoded logic for any provider. 6 V2 manifests with MCP/CU/Multimodal declarations.

Global Providers

OpenAI OpenAI
Anthropic Anthropic
Google Gemini Google Gemini
Groq Groq
Mistral Mistral
Cohere Cohere
Azure OpenAI Azure OpenAI
Together AI Together AI
Hugging Face Hugging Face
Grok Grok
Ollama Ollama
D
DeepInfra
O
OpenRouter
N
NVIDIA
F
Fireworks
R
Replicate
A
AI21 Labs
C
Cerebras
P
Perplexity
L
Lepton AI

China Region

DeepSeek DeepSeek
Qwen Qwen
D
Doubao
Z
Zhipu GLM
Baidu ERNIE Baidu ERNIE
Tencent Hunyuan Tencent Hunyuan
iFlytek Spark iFlytek Spark
Moonshot Moonshot
M
MiniMax
B
Baichuan
Y
Yi / 01.AI
S
SiliconFlow
S
SenseNova
T
Tiangong

Ready to get started?

Read the documentation, pick your runtime (Rust, Python, TypeScript, or Go), and start building with 37+ AI providers today.

Quick Start Guide Read the Docs