nasiko

Nasiko is a comprehensive control plane designed to solve the operational challenges of managing multiple AI agents. By acting as a central proxy, it ensures that agents are never directly exposed to the public internet. Every inter-agent call is routed through the Nasiko server, which enforces rate limits, access control lists (ACLs), and provides end-to-end observability via OpenTelemetry. Key capabilities include: - **A2A Protocol Support**: Deploys and manages any agent that adheres to the A2A specification, regardless of the underlying programming language (Python, Rust, Go, TypeScript). - **Intelligent Routing**: Features a three-stage pipeline (shortlist, rerank, and LLM selection) to route requests to the appropriate agent automatically. - **Security & Secrets**: Provides AES-256-GCM encrypted secret storage and an LLM router that replaces raw API keys with short-lived identity tokens, preventing credential leakage. - **MCP Gateway**: Integrates with Model Context Protocol (MCP) servers to provide agents with permission-filtered tool access. - **Observability**: Automatically collects token usage, costs, and trace data, with built-in support for OTel, Tempo, and Loki. - **Developer Experience**: Includes a CLI for managing agent lifecycles, building/pushing images to an embedded OCI registry, and configuring coding-agent integrations (e.g., Claude Code, Cursor, OpenCode). The architecture is built as a single process that leverages Postgres, Redis, and S3 for state management, making it suitable for self-hosted environments via Docker Compose.