Platos

Platos is an open-source, self-hostable agent runtime designed to support the deployment, monitoring, and management of AI agents in production environments. It provides a unified platform where teams can define agents with system prompts, tools, memory, and version pins, and operate them via REST, WebSocket, or through the Model Context Protocol (MCP). The platform is intended for organizations and developers seeking to own their infrastructure and data while maintaining flexibility in model and tool choices.

Key features of Platos include the ability to federate backends through an MCP gateway, support for bring-your-own-provider (including Anthropic, OpenAI, Google Vertex, and Voyage), and multi-agent clusters that enable agents to share user memory and thread history. Its memory subsystem offers conversation history, semantic recall, knowledge graphs, and user profiles, all scoped per organization, project, environment, and user, with encryption at rest. Platos also incorporates durable background operations, approval workflows for sensitive tool calls, a skills framework for reusable agent behaviors, and full observability with tracing, span timelines, and cost attribution per agent, provider, and token. Observability data can be exported to ClickHouse, and the system supports cost tracking on every provider key.

The platform is delivered as a self-hosted solution that can be deployed locally in minutes using Docker Compose. It is structured in four architectural layers: clients (including a web dashboard and chat), the runtime and MCP gateway, integrated tools and memory, and foundational data services such as Postgres, ClickHouse, Redis, and MinIO. dev. Platos supports operator approval workflows, per-tool access controls, and per-agent budget caps. 0 and is developed by Winsen Labs.

Platos addresses the need for an open, ownable, and production-ready agent runtime stack, enabling teams to avoid the trade-offs of managed platforms or the overhead of building their own infrastructure from scratch.