Iris
Routes AI prompts across multiple providers, intelligently selecting the most cost-effective or preferred model and learning over time.
Iris is a self-hosted, multi-provider AI router designed to optimize LLM interactions by dynamically selecting the cheapest-acceptable model for each task. Unlike library-based or hosted solutions, Iris runs as a local service, persisting provider statistics and conversation histories to learn and bias future routing decisions. It operates with a strong focus on privacy, keeping all data local and providing an integrated MCP server for seamless client interaction without external telemetry.
Key Features
01Council fan-out for parallel prompting across N providers with side-by-side responses and optional LLM-based judging.
02Persistent learning loop that biases model selection based on historical success rates and user preferences via localmem.
03Robust features including per-provider circuit breakers, native tool/function calling, Anthropic prompt caching, and structured outputs.
04Multi-provider AI routing supporting 16 built-in (5 native, 11 OpenAI-compatible) and custom APIs.
05Integrated Model Context Protocol (MCP) server offering 13 tools, Prometheus metrics, mDNS advertisement, and live event streams.
062 GitHub stars
Use Cases
01Managing persistent AI conversation sessions and context across multiple command-line invocations.
02Deploying a self-hosted, intelligent AI gateway for internal applications with strong privacy and local-first preferences.
03Optimizing AI API costs and performance by intelligently routing requests to the best available LLM provider.
04Benchmarking and comparing multiple LLM responses in parallel for complex queries using the council fan-out feature.