GitLab 19.5 release notes

GitLab released version 19.5 on October 15, 2026, making the GitLab MCP server and GitLab Secrets Manager generally available. The MCP server gives MCP-compatible agents such as Claude Code and Cursor direct access to GitLab's expanded tool catalog and GitLab Duo Agent Platform agents and flows, while Secrets Manager, built on OpenBao, scopes each CI secret to the job that needs it and generates audit logs for every create, read, update, and delete. The release also extends the AI Governance Dashboard to Externally Connected Agents like Claude Code, Cursor, and Codex via a one-time `glab govern setup`, in beta, and makes non-default branch vulnerability tracking generally available.

On October 15, 2026, GitLab 19.5 was released with the following features. We are excited to recognize Dhairya https://gitlab.com/DhairyaMajmudar , a Level 3 contributor, as this month’s Notable Contributor https://contributors.gitlab.com/notable-contributors Dhairya contributed to offline transfer and Organizations support for import, so that groups, projects, and their memberships can move between GitLab instances. In previous versions of GitLab, MCP clients could reach only a limited set of GitLab tools. Now the GitLab MCP server is generally available. With the MCP server, any MCP-compatible agent, such as Claude Code or Cursor, gets direct access to the expanded GitLab catalog of tools and GitLab Duo Agent Platform agents and flows. Your agents can work across more of your GitLab workflow and run Agent Platform agents and flows without leaving your preferred tool. Without a centralized secrets store, credentials can become scattered across over-scoped CI/CD variables, local config files, and standalone vaults. This makes them harder to manage, protect, and audit. GitLab Secrets Manager is now generally available on GitLab.com and GitLab Self-Managed. Built on OpenBao, it uses your existing group and project permissions, so there’s no separate access model to maintain. Each CI secret is scoped to the job that needs it, based on environment, branch, and branch protection rules. Audit logs are generated for every create, read, update, and delete. GitLab Secrets Manager consumes GitLab Credits. Start a free 30-day trial https://docs.gitlab.com/ci/secrets/secrets manager/credit usage/ start-a-trial or enable it for your group or instance https://docs.gitlab.com/ci/secrets/secrets manager/ enable-gitlab-secrets-manager . Previously, the AI Governance Dashboard https://docs.gitlab.com/user/ai-governance/governance-dashboard/ showed activity only from GitLab Duo Agent Platform agents. This provided you total visibility into all of your Agent Platform agents, including the sessions they were running, including tool calls, MCPs used, and the prompt input and output from the user and agent respectively. With the GitLab 19.5 release, after a one-time glab govern setup , you can now extend governance visibility past just GitLab Duo Agent Platform agents, and connect Externally Connected Agents ECAs like Claude Code, Cursor, and Codex to GitLab AI Governance by using the GitLab CLI glab . When Allow external agents and Store AI audit events are enabled, security and compliance teams get visibility into ECAs that run outside GitLab, alongside your GitLab Duo Agent Platform agents. Every qualifying agent session is captured, including who ran it, which agent, which project, and when, with sessions and tool calls appearing in the AI Governance Dashboard and the AI audit events report. Audit events also flow to your SIEM through audit event streaming. Tool output is not sent, except for short error messages from failed tool calls. This feature is in beta. Share your feedback in issue 633489 https://gitlab.com/gitlab-org/gitlab/-/work items/633489 . Non-default branch tracking is now generally available. You can track vulnerabilities on branches other than the default branch. For the best results, target a small number of long-lived release branches, such as branches for specific environments project-qa , project-prod or deployment platforms project-iOS , project-android . This release includes the following capabilities: Administrators can now select any subgroup, not just top-level groups, in the Restrict access based on group membership https://docs.gitlab.com/administration/gitlab duo/configure/access control/ setting for GitLab Duo. You can scope GitLab Duo Agent Platform access to a specific team or pilot group without restructuring your group hierarchy. Previously, this setting accepted only top-level groups. Organizations with a single top-level group for all users had no supported way to limit GitLab Duo Agent Platform access to a subset of users. Now, when you select a subgroup, direct members of that subgroup get access. This makes it easier to run a GitLab Duo Agent Platform pilot or roll out access one division at a time on GitLab Self-Managed and GitLab Dedicated. The /goal slash command in GitLab Duo CLI is now generally available in GitLab Duo CLI 9.27.0. Describe an open-ended goal, and GitLab Duo delegates it to a governed, goal-driven flow that runs locally and handles both implementation and verification. To start, run /goal <task . For example: /goal Fix the failing tests in spec/models/user spec.rb In GitLab 19.5 GitLab Duo CLI 9.24.0 , the Run suggested prompts as /goal sessions setting in /settings runs suggested prompts as /goal sessions instead of chat messages. /goal requires GitLab 19.3 and later, and GitLab Duo CLI 9.17.0 and later. GitLab Duo CLI has a new auto mode. In auto mode, GitLab Duo uses tools without asking for approval first, so longer tasks run without stopping for each shell command, git operation, or MCP tool call. Auto mode respects your organization’s controls: To use auto mode, you need GitLab Duo CLI 9.22.0 or later. Press Tab until the mode under the prompt shows auto . Use auto mode only in repositories you trust and on branches you can discard. You can now configure GitLab Duo CLI with settings files, for yourself or for your whole organization. The settings file stores the choices you make in the /settings panel. You can also edit the file directly, or create it in advance, for example in a container image. System administrators can deploy a managed settings file to each machine with a mobile device management tool or group policy. Settings in the managed file take precedence over the settings file, command-line options, and environment variables, and users can’t change them. Settings not in the managed file remain under user control. The settings file requires GitLab Duo CLI 9.26.0 or later, and the managed settings file requires GitLab Duo CLI 9.27.0 or later. In previous versions of GitLab, when you asked GitLab Duo to review a merge request again, it treated the review as a first pass. It could repeat findings from its earlier threads, even ones you had already addressed. Now, Code Review Flow uses comment threads from earlier GitLab Duo reviews, including your replies, to evaluate new changes. GitLab Duo focuses on lines added after its last review, does not repeat points it already raised, and lists outstanding findings in its summary. If you reply on a thread to explain why the code is correct, GitLab Duo does not raise the point again. To start a follow-up review, ask GitLab Duo to review the merge request again. With Start a new review on push turned on, GitLab Duo automatically starts a follow-up review when you push changes. Previously, you could not control which AWS credentials the AI Gateway used for your Amazon Bedrock models. A single AWS identity needed access to all these models. With this release, you can specify an IAM role for each model, and the AI Gateway assumes that role when calling the model. This way, each model gets only the AWS access it needs, even when the role is in another AWS account. Previously, GitLab Duo CLI could read text files but could not use screenshots or diagrams as context. Now, when you run the CLI locally in interactive or headless mode, it can read images from your repository and attachments on issues, merge requests, and comments in the same project. Image context requires GitLab Duo CLI 9.27.0 or later and a GitLab-managed model. GitLab Duo CLI now helps new users get started without leaving the terminal: /goal progress, and MCP status. /whatsnew command point you to new features. To get the full experience, update to GitLab Duo CLI 9.28.0 or later. Previously, if you needed an agent to check something inside a CI/CD job’s build artifacts, it had to download and unzip the whole archive just to read one file. Now the get artifact file MCP tool lets your agent read a single file directly from a job’s artifacts, whether it’s a log, a generated report, or a test fixture. New GitLab Duo agent and flow tools let you invoke and track GitLab Duo agents and flows from any MCP client: start duo session starts a session. get duo session reads a session’s current state. list duo sessions pages through your session history. list duo agents and flows lists the GitLab Duo agents and flows enabled in a project. Previously, an agent connecting to GitLab over MCP had no way to discover which groups and subgroups you have access to without you hardcoding IDs or paths. The new list groups tool lets your agent browse your organization’s group hierarchy on its own, so it can find the right group before acting instead of guessing at identifiers. You can now choose which toolsets the GitLab MCP server exposes. Every additional tool that isn’t required wastes space in the LLM’s context window, and runs the risk of confusing the LLM, especially if you are using multiple MCP servers at once. The redesigned GitLab Duo Agentic Chat UI in VS Code and JetBrains IDEs, introduced as a beta in GitLab 19.3, is now generally available and is the default layout. You no longer need to turn on beta and experimental features to use it. The new layout has no tabs. You start conversations, view your chat history, and select an agent and model in the same view. The new layout is the default in: To use the Software Development Flow or view the flows you have run, switch to the classic layout from the panel. If you connect an agent to GitLab over MCP today, nothing stops it from reconnecting or retrying faster than your workflow actually needs, and on GitLab.com that carries a cost. Starting October 15, the GitLab MCP server on GitLab.com limits each user to 60 requests a minute on the Free tier and 600 on Premium and Ultimate. Every reconnect and retry counts toward your limit. If your agent starts getting rate-limited HTTP 429 errors, check its reconnect and retry settings before upgrading to a higher tier. In previous versions of GitLab, you had to check each approval rule and guess who had time to review your merge request. Or you assigned a whole group and hoped someone picked it up. Now you can use the Recommend Reviewers Flow to suggest reviewers instead. The flow is generally available. For each approval rule, it recommends the minimum number of reviewers needed, based on availability, workload, time zone, and recent activity. Recommendations appear in the Recommended section of the merge request sidebar. Hover over a user to see why they were recommended. To get started, turn on the flow for your top-level group. It runs when a merge request is created, or you can create a trigger to run it when a draft is marked ready. To get new recommendations, select Refresh . The Software Development Flow now runs on the GitLab Duo Agent Platform flow registry, the same foundation as other flows. The flow works the same way in your IDE, and new fixes and improvements go to this version. To keep using the Software Development Flow, update your editor extension to one of these versions or later: In previous versions of GitLab, the API endpoint you use to trigger GitLab Duo Agent Platform flows from CI/CD pipelines, scripts, and other automation was an experiment, so it was hard to know whether you could rely on it in production. Now the trigger a flow endpoint https://docs.gitlab.com/api/duo agent platform flows/ trigger-a-flow of the Flows API is generally available, so you can build production automation on it knowing it follows the GitLab REST API breaking changes policy. Many development teams split large CI/CD setups into parent and child pipelines so each part is easier to manage. Until now, a child pipeline job couldn’t safely depend on a parent job that only runs under certain conditions, such as tests that run only when certain files change. If that parent job was skipped, the child job broke, so teams had to work around it with duplicate jobs or brittle scripts. Now you can mark these dependencies as optional. If the parent job runs, the child waits for it as usual. If it doesn’t run, the child job goes ahead without it. Pipelines stay fast and reliable without the workarounds, and a failure in a job that does run still stops the pipeline, so quality checks aren’t weakened. Thank you to Jimmy Spagnola https://gitlab.com/jspagnola for this contribution You can now customize Secret Detection and SAST profiles in the UI. Set scan options like excluded paths, custom analyzer images, and full Git history scans once for a group, and apply them to all its projects. Previously, a security configuration profile enabled a scanner with its default settings. Anything beyond the defaults required edits to .gitlab-ci.yml in each project. Now the profile includes those settings too. Currently, customization of Dependency Scanning profiles is only available through the GraphQL API. UI customization is planned in a future release. In previous versions of GitLab, permalinks https://docs.gitlab.com/user/project/repository/files/ create-permalinks to a file’s lines rendered as plain URLs, so you had to leave the merge request, issue, or wiki page to view the referenced code. Now, permalinks render as a syntax-highlighted snippet with the file path and commit, so you can read the code without leaving the page. You can only view embedded snippets if you have read access to the repository. In previous versions of GitLab, Observability was an experiment. Now it’s in beta on GitLab.com and GitLab Self-Managed. With GitLab Observability, you can send traces, metrics, and logs with OpenTelemetry, then explore them next to your code, merge requests, and CI/CD pipelines. On GitLab.com, each group or personal namespace opts in by selecting Enable Observability . On GitLab Self-Managed, you host the Observability backend and connect it to a group or personal namespace. Observability is available at no cost during the beta. In previous versions of GitLab, to find information in a large Markdown table, you had to scan every row or copy the table into a spreadsheet. Now you can sort rendered Markdown tables by any column, in both view mode and preview. In a column header, select the sort button to sort rows in ascending or descending order, then reset to restore the original order. You can also sort with a keyboard. In previous versions of GitLab, when you scrolled through a long table, the header row scrolled out of view, so it was hard to tell what each column contained. Now, table headers stay visible at the top of the table as you scroll, in view mode, preview, and the rich text editor. Sticky headers are always on, so you don’t need to change any settings. Business Logic Security Analyzer is a GitLab Duo Agent Platform foundational flow that reasons about what your code is meant to enforce. It finds vulnerabilities that pattern-based scanners often miss: broken object-level authorization, missing authorization checks, and mass assignment. Start a whole-repository scan from GitLab Duo Chat or the API. To scan merge requests automatically, attach the Business Logic scan profile to a project. Findings appear in the vulnerability report and the merge request security widget. Business Logic Security Analyzer is in closed beta for GitLab Ultimate customers on GitLab.com. To request access, contact your account team. Each scan consumes GitLab Credits. For prerequisites and setup, see Business logic scanning https://docs.gitlab.com/user/application security/business logic scanning/ . GitLab Relay KAS now uses gRPC by default. GitLab.com advertised grpcs://kas.gitlab.com in GitLab 19.4. In GitLab 19.5, new Linux package and Helm chart installations advertise grpcs:// too. New agent installations connect over grpcs:// instead of gRPC wrapped in WebSocket, which means one protocol to route and troubleshoot, and less overhead for each connection. Agents that already connect over wss:// work, but you should move them. On GitLab.com, set the agent address to grpcs://kas.gitlab.com . For example, use --set config.kasAddress=grpcs://kas.gitlab.com when you upgrade the gitlab-agent Helm chart. On GitLab Self-Managed, use the address shown in the agent installation command https://docs.gitlab.com/user/clusters/agent/install/ after you upgrade. You can now share a filtered view of the security inventory with a link. Filter by tool coverage, vulnerability count, or project name, then copy the URL and send it. Whoever opens the link sees the same view. Previously, only the project name search was saved in the URL, so a shared link lost the other filters. Now the full filter state is preserved, so you can bookmark a view or hand it to the team that owns those projects. You can now start a custom flow directly from GitLab Duo Agentic Chat, without needing to configure a trigger first. Previously, running a custom flow required setting up a trigger event — a mention, an assignment, or similar — which made flows difficult to test and iterate on. Now, you can launch any AI Catalog flow enabled in your project by sending a message in GitLab Duo Agentic Chat. Your message becomes the goal for the flow. If you send no goal, the flow falls back to its own description. Chat links to the resulting agent session so you can follow along. If you used GitLab Credits, you had no way to check your own usage. Only group Owners and administrators could open the GitLab Credits dashboard, so you had to ask someone with access. You can now see your own usage in User settings GitLab Credits . The page shows your total usage, daily average, and peak day, and you can filter by product or select a date range. If you belong to more than one group with GitLab Credits, you can switch between them. Administrators can now resync a single LDAP-blocked user from the Admin area Users page. The new Resync with LDAP action runs an LDAP access check and unblocks the user when LDAP allows access, or leaves the user blocked when it does not. If the user has no LDAP identity, the action declines with a message. In previous versions of GitLab, an LDAP-blocked account stayed blocked until the daily user sync ran or the user signed in, which service accounts cannot do. Thank you to Sergey Pechenko https://gitlab.com/tnt4brain for this contribution MR https://gitlab.com/gitlab-org/gitlab/-/merge requests/255974 You can now cancel an active auto-merge with the cancel auto merge API endpoint. If the merge request is on a merge train, the call also removes it from the train. The older cancel merge when pipeline succeeds endpoint is now deprecated. Separately, git push -o merge request.auto merge is now rejected with a remote message when you cannot merge to the target branch. In previous versions of GitLab, the auto-merge was recorded and left the merge request unmergeable until canceled. Thank you to the following users for these contributions Administrators can now set how often job logs update, under Admin area Settings CI/CD or through the Settings API. You can select between 3 to 3600 seconds. In previous versions of GitLab, these intervals were constants, so you could not shorten the initial live-log delay for short jobs without patching the instance. Thank you to Aleksandr Kotlyar https://gitlab.com/aleksandr-kotlyar for this contribution MR https://gitlab.com/gitlab-org/gitlab/-/merge requests/251401 GraphQL now exposes Repository.tags , with forward pagination, search, and a TagSort enum for name, updated, and version order, at up to 100 tags per page. Repository.tags is an experiment in GitLab 19.5. TerraformState.versions lists a state’s versions, newest first. The merge requests REST API now returns author.bot on single and list responses, so a client can tell bot and service account authors apart without a second call to the Users API. You can now use %{commits count} in merge commit and squash commit message templates, and in new merge request description templates. For example, %{title} %{commits count} commits renders as Bugfix 3 commits . Thank you to Jeston Singh https://gitlab.com/sjestonsingh for this contribution MR https://gitlab.com/gitlab-org/gitlab/-/merge requests/255231 When you submit a review on a merge request you are not yet a reviewer of, with Comment , Approve , or Request changes , GitLab now adds you to the Reviewers list. The review drawer tells you beforehand: “You will be added as a reviewer.” The Reviewers list and the people who took part in the review no longer drift apart. This behavior has been available since GitLab 19.4, when its feature flag was removed. Thank you to Michael Fischer https://gitlab.com/mfischer5 for this contribution MR https://gitlab.com/gitlab-org/gitlab/-/merge requests/250880 GitLab now ignores user IDs that GPG reports as revoked. A revoked identity no longer appears under Access GPG keys in user settings and no longer counts as a verified email for the Verified badge on signed commits. Commits signed under a revoked identity move to “same user, different email” or “other user”, and a key whose every identity is revoked verifies nothing. Signatures already verified under a key that stays in place are not re-derived. Thank you to José M. Requena Plens https://gitlab.com/jmrp for this contribution MR https://gitlab.com/gitlab-org/gitlab/-/merge requests/255300 Other improvements in this release include: add branch returns the branch web url , tool aliases in the enabled-tools header resolve, mismatched URL and project ID inputs are rejected, and ai workflows tokens can read blobs in batch. CI JOB TAGS and CI RUNNER TAGS JSON-encode each tag. Duplicate variable keys in push options are rejected. ?scopes= links prefill the token form. Other additions to the interface in this release include: .4dm files are syntax highlighted. Approval rules, compliance frameworks, and protected environments require Premium or Ultimate.