kevinelliott/agentpipe

Built in Go, AgentPipe uses a standardized three-part prompt structure and stdio transport to enable seamless interoperability between heterogeneous AI CLI tools, supporting everything from local models (Ollama) to API-based providers (OpenRouter). The system implements production-grade reliability through rate limiting with token buckets, exponential backoff retry logic, and middleware-based message processing, while offering real-time observability via Prometheus metrics tracking tokens, costs, and agent health. Multiple conversation modes (round-robin, reactive, free-form) combined with persistent state management, export formats, and an interactive TUI with agent filtering and search create flexible orchestration for complex multi-agent workflows.