GitLab 19.5 puts external code agents in its audit controls

GitLab released version 19.5 on 15 October. Its MCP server gives AI agents access to GitLab through the Model Context Protocol, a standard for agents. It is open to all users. Tools such as Claude Code, Cursor and Codex can connect to it. Safety teams can see which agents operate, who uses them and on which projects.

What is new

Administrators select which toolsets an agent can use. This cuts the number of tools and the context that the agent uses.

GitLab records the user, the agent, the project and the time of each session. It can send audit records to a tool for safety checks. The Duo CLI gets an auto mode, off at first, where agents do commands with no approval for each one. Tools with the name Always Deny stay blocked.

The numbers

On the GitLab cloud, MCP use has a limit of 60 requests each minute for Free users and 600 for Premium and Ultimate users. GitLab gave no test results.

What is not known

The source does not tell how many companies use the controls. It tells that GitLab follows GitHub in its work to make agents part of the software cycle. It does not compare the methods.

Insight: GitLab wants to be the gate for agents it did not build

In October, Descope released a hub that gives each agent an owner and an approval step. Slack opened its Code API to all developers. GitLab, Descope and Slack each read as a move to be the gate for agents from other companies. The first gate that companies use can set the standard. Look at the gate that your safety team uses before you add one more.