GitLab 19.5 makes the GitLab MCP server generally available. MCP-compatible tools such as Claude Code and Cursor can access GitLab's broader tool catalogue and start or monitor GitLab Duo Agent Platform agents and flows. Administrators can select which MCP toolsets are exposed, which reduces unused tools and the context they consume. GitLab.com also sets MCP rate limits of 60 requests per minute for Free users and 600 for Premium and Ultimate users, with reconnects and retries counted.
Externally Connected Agents can now be brought into GitLab's AI Governance capabilities. Claude Code, Cursor and Codex can be connected through the GitLab CLI, and qualifying sessions record the user, agent, project and timing alongside GitLab's own agents. Audit events can be streamed to a SIEM. GitLab Duo CLI also adds an auto mode, disabled by default, that lets agents run tools without approval for each shell command, Git operation or MCP call, while tools marked Always Deny stay blocked.
GitLab Secrets Manager is now generally available. Built on OpenBao, it uses existing group and project permissions, scopes secrets to the jobs that need them and generates audit events for secret operations. Security configuration profiles now let organisations centrally set excluded paths, analyzer images and full-history scans for Secret Detection and SAST.
Why it matters if you build agents
Agents running in external tools can now reach GitLab repositories, issues, pipelines and artifacts through MCP, so builders should plan for per-tier rate limits of 60 or 600 requests per minute. Toolset selection gives a way to limit which actions an agent can take. If your agents connect through the GitLab CLI, their sessions can be audited, and auto mode should stay off unless the instance, group or project settings allow it.
Get stories like this every weekday
One short email with the day's agent tooling news.