🦊GitLab 19.5 Puts AI Agents Under Governance With 600 MCP Requests Per Minute

Your AI coding agents now need a governance policy

GitLab just made agentic development a governed part of its DevOps platform, and that's a bigger deal than the feature list makes it sound. Version 19.5 lets Claude Code, Cursor, and Codex connect through the GitLab CLI, and once they're configured, your security team can see which agent touched which project, who ran it, and when. If your team lets AI agents near repos, pipelines, or credentials, this is the release that stops it from being the Wild West. The new MCP server gives agents access to groups, CI/CD artifacts, and Duo agent flows, but admins can now choose exactly which toolsets are exposed. Fewer tools means less context burned and fewer ways for an agent to go sideways. The numbers matter here. GitLab.com caps MCP requests at 60 per minute for Free users and 600 for Premium and Ultimate, and that limit counts reconnects and retries, so a flaky connection can eat your quota fast. Audit events stream to a SIEM, and the new Secrets Manager, built on OpenBao, respects existing group and project permissions. Watch the rate limits if you run chatty agents in CI.