BlockEdge

Technology and digital life

GitLab 19.5 expands agent access and governance controls

GitLab 19.5, released on October 15, 2026, made its Model Context Protocol (MCP) server generally available and expanded how external AI agents can interact with the DevOps platform. This is a retrospective account based on an InfoQ report published November 15, 2026.

The MCP server lets compatible tools, including Claude Code and Cursor, access GitLab capabilities. New tools can discover groups, inspect individual CI/CD artifact files, and start or monitor GitLab Duo agents and workflows. Administrators can choose which MCP toolsets to expose, limiting the tools available to an agent and the context they consume. On GitLab.com, MCP requests are limited to 60 per minute for Free users and 600 per minute for Premium and Ultimate users; reconnection and retry requests count toward those limits.

GitLab 19.5 also extends AI governance beyond GitLab Duo agents. Organizations can connect external tools such as Claude Code, Cursor and Codex through the GitLab CLI. Once configured, they can record eligible agent sessions with information about the user, agent, project and time, and view that activity alongside GitLab’s own agents. Audit events can also be streamed to a SIEM system.

For GitLab Duo CLI, a new auto mode allows tool execution without approval for each shell command, Git operation or MCP call. It is off by default, can be controlled at instance, group and project levels, and does not override tools marked “Always Deny.” The /goal command is also generally available, letting developers describe an open-ended objective for GitLab Duo to implement and verify through a governed workflow.

The release additionally makes GitLab Secrets Manager generally available. Built on OpenBao, it integrates with GitLab group and project permissions, scopes secrets to the tasks that need them, and generates audit events for secret operations. Expanded security configuration templates let organizations centrally set parameters for Secret Detection and static application security testing, including excluded paths, analyzer images and full-history scans, rather than repeating them in each project’s CI/CD configuration.

Source: infoq.cn (originally published 2026-11-15 09:00 UTC+8)