Harnesses

GitLab 19.5 integrates external AI agents into its governance framework

November 17, 2026 · 3 min read

Also published in Deutsch

A laptop screen displaying code and a steaming mug on a desk
Mohammad Rahmani / Unsplash

GitLab released version 19.5 on October 15, 2026, positioning agent software development and governance as core platform capabilities. The update makes the GitLab MCP server generally available and expands how external AI agents interact with the platform. This release allows compatible tools to access a broader directory of GitLab tools and invoke agents and workflows from the GitLab Duo Agent Platform. Simultaneously, the version introduces governance controls for external agents, enabling security teams to identify which agents are in use, who is using them, and which projects they target.

The generally available GitLab MCP server opens platform capabilities to agents via the Model Context Protocol. New tools allow agents to discover groups, inspect individual CI/CD artifact files, and start or monitor GitLab Duo agents and workflows. Administrators can select which MCP tool sets to expose, reducing unnecessary tools and their associated context consumption. Providing agents with all available tools increases context usage and the number of potential model actions. Tool set filtering serves as a lightweight least-privilege proxy design. GitLab.com also introduced MCP rate limits, allowing free users 60 requests per minute and Premium or Ultimate users 600 requests per minute. Reconnection and retry requests count toward these limits.

A significant addition is the inclusion of externally connected agents within GitLab's AI governance scope. Previously, AI governance dashboards focused primarily on GitLab Duo Agent Platform agents. Version 19.5 allows tools such as Claude Code, Cursor, and Codex to connect via the GitLab CLI. Once configured, organizations can log qualifying agent sessions, including user, agent, project, and time information. This activity displays alongside GitLab's own agents, and audit events can stream to SIEM systems. This approach moves beyond providing an MCP endpoint, establishing that connectivity provides capability while governance determines safe usage at an organizational scale.

The GitLab Duo CLI gained an automatic mode, allowing agents to execute tools without requesting approval for every shell command, Git operation, or MCP call. This feature is disabled by default and controllable at the instance, group, or project level. Tools explicitly marked as Always Deny remain blocked. The /goal command also became generally available, allowing developers to describe an open-ended goal for GitLab Duo to complete code implementation and result verification through governed processes. These features drive GitLab toward an agent-driven development model where developers express intent and agents execute multi-step tasks automatically. As agent autonomy increases, relying on manual approval for each operation becomes less practical, making governance controls increasingly important.

The release also officially launched GitLab Secrets Manager, built on OpenBao. It integrates with existing GitLab group and project permissions, limiting secret scope to specific tasks and generating audit events for secret operations. GitLab expanded security configuration template capabilities, allowing enterprises to uniformly configure parameters for Secret Detection and Static Application Security Testing, such as exclusion paths and analyzer images, without repeating configurations in every project's CI/CD file. These capabilities strengthen control over codebases, credentials, and pipelines accessible to agents, complementing the new agent features. The platform's direction aligns with making code agents participants in the software lifecycle, focusing on integrating external agents into its existing DevSecOps governance system rather than just providing coding assistance.