ai.matey.backend

Server-side backend provider adapters for AI Matey - Universal AI Adapter System.

Part of the ai.matey monorepo.

Installation

npm install ai.matey.backend

Included Providers

This package includes adapters for 30 AI providers:

Commercial APIs

• OpenAI - GPT-5.6 family
• Anthropic - Claude Sonnet 5, Claude Opus 4.7+
• Google Gemini - Gemini 3.6 Flash and other current-generation Gemini models
• Mistral AI - Mistral Large, Medium, Small
• Cohere - Command, Command-Light, Command-R
• xAI - Grok models
• AI21 Labs - Jurassic models
• Moonshot AI - Kimi models with long-context support (up to 128K)
• Inception Labs - Mercury diffusion language models
• Alibaba Cloud Model Studio (DashScope) - Qwen model family, OpenAI-compatible mode

Cloud Providers

• AWS Bedrock - Amazon's managed AI service
• Azure OpenAI - Microsoft's OpenAI deployment
• Cloudflare Workers AI - Edge AI deployment

Fast Inference

• Groq - Ultra-fast LLaMA, Mixtral inference
• Fireworks AI - Fast inference platform
• Together AI - Open model hosting
• Anyscale - Fast endpoints
• DeepInfra - High-performance inference
• Cerebras - AI supercomputer inference
• SambaNova - High-throughput RDU-accelerated inference

Aggregators

• OpenRouter - Multi-provider routing and fallback
• Perplexity - Search-augmented models
• GitHub Models - Free access to OpenAI, Meta, DeepSeek, Mistral & Microsoft models via any GitHub account

Specialized

• Replicate - ML model deployment
• NVIDIA NIM - NVIDIA inference microservices
• Hugging Face - Open model inference
• DeepSeek - Research models

Local/Development

• Ollama - Local model hosting
• LM Studio - Local desktop inference
• OmniRoute - Self-hosted gateway fronting 290+ providers (90+ free), no API key required by default

For browser-compatible adapters (Chrome AI, Function, Mock), see ai.matey.backend.browser.

Usage

import { OpenAIBackendAdapter, AnthropicBackendAdapter } from 'ai.matey.backend';

// Create an OpenAI backend
const openaiBackend = new OpenAIBackendAdapter({
  apiKey: process.env.OPENAI_API_KEY,
});

// Create an Anthropic backend
const anthropicBackend = new AnthropicBackendAdapter({
  apiKey: process.env.ANTHROPIC_API_KEY,
});

Subpath Imports

You can also import specific providers directly:

import { OpenAIBackendAdapter } from 'ai.matey.backend/openai';
import { AnthropicBackendAdapter } from 'ai.matey.backend/anthropic';

Structured Output

Set responseFormat on an IRChatRequest to get schema-constrained JSON output. OpenAI, Anthropic, and Gemini map it to their native structured-output mechanisms (response_format/output_config/responseSchema); every other backend falls back to prompt injection + best-effort JSON extraction. See docs/IR-FORMAT.md for the full support matrix and a request/response example.

Anthropic Sampling Parameters

Claude Opus 4.7+ and Sonnet 5 return an HTTP 400 if temperature/top_p/top_k are set to non-default values. AnthropicBackendAdapter detects these model families and omits the params automatically - no config needed, but don't rely on sampling-param overrides taking effect against these specific models.

API Reference

See the TypeScript definitions for detailed API documentation.

License

MIT - see LICENSE for details.