CoreWeave wires coding agents into GPU infrastructure with a Mission Control MCP server

CoreWeave introduced its Mission Control MCP server on September 24. The hosted server lets MCP-compatible coding agents investigate CoreWeave infrastructure from a developer’s existing workspace. CoreWeave names Claude Code, Cursor, and Codex as supported clients.

What the server exposes

According to CoreWeave’s setup documentation, users connect a client to the hosted endpoint and authenticate with a CoreWeave API token. The available tools cover metrics, logs, dashboards, cluster and node state, SUNK resources, and object storage context. The agent’s visibility follows the permissions attached to that token. The announcement describes the hosted MCP server as read-only. Its documentation also describes limited write tools that CoreWeave can enable for an organization. Teams should check their own tool permissions before treating an agent session as read-only. The tool catalog exposes individual operations that an engineer can inspect during an investigation. It also makes CoreWeave’s public documentation searchable from the client. For a team investigating a slow training job, the practical gain is access to infrastructure evidence in the coding client already open on the engineer’s desk. The release does not show that an agent can diagnose every incident or act without approval; the useful next test is whether its evidence and permissions hold up during a real investigation.