MCP-Universe RL trains tool-use agents by using MCP servers as the RL environment interface

arXiv·medium signalXBlueskyLinkedInCopy link

Most RL frameworks stop at the policy update and leave the user to stand up isolated environments for hundreds of concurrent trajectories and to keep the GPU busy through long episodes stalled on slow tool calls. MCP-U RL uses the Model Context Protocol as the environment interface, so any existing MCP server plugs into training with no RL-specific integration code, and adds an environment-orchestration layer over a pluggable container backend plus a staged rollout pipeline that overlaps trajectories. With one configuration changed only by task spec, the authors trained software-engineering, deep-research and general tool-use agents on gpt-oss-20b, with veRL and slime backend integrations.