ToolVerse Builds an Agentic RL Environment From ~400 Real MCP Servers and ~4,500 Tools

arXiv·medium signal

ToolVerse (arXiv 2607.15660, July 17, Shuaiyu Zhou et al.) tackles the shortage of realistic long-horizon training environments for agentic RL by auto-constructing environments from roughly 400 real-world Model Context Protocol servers exposing about 4,500 tools. It generates the GUST task set via a Dynamic Unlocking Sampling Algorithm over tool-dependency graphs, and introduces Turn-Aware Relative Advantage to fix credit assignment across long trajectories. Notable as one of the first papers to treat the deployed MCP ecosystem itself as RL training infrastructure.