GitLab wires Claude Code into Duo Agent Platform with a skill
A walkthrough shows one installable skill that configures a whole GitLab group so issues, reviews and pipeline fixes run on GitLab while developers stay in their terminal.
Setup lives inside the harness
On November 18, 2026, GitLab published a walkthrough of agentic engineering with GitLab and a local agent that centres on a GitLab Duo Agent Platform setup run from the terminal. GitLab uses Claude Code as its example because that is what it says most customers use, and says the same flow works with Codex, Cursor, OpenCode, Duo CLI and other harnesses.
The mechanics are small: clone the agentic-eng-config repository, copy the gitlab-duo-setup and gitlab-mr-followup skills into ~/.claude/skills/, and run /gitlab-duo-setup. The post notes that other harnesses may invoke skills with different syntax. Longer, repository-aware work then runs on GitLab.
What the skill changes
The skill reads the top-level group and its projects first and shows one plan listing every change and whether it lands on a group or a project. It waits for a yes, and any line can be dropped. Then it enables Duo Agent Platform and flow execution for the group, adds the Code Review, Duo Developer and Fix CI/CD Pipeline flows while leaving existing flows on, turns on automatic code review for the group, and creates an automation in every project with CI/CD that runs Fix CI/CD Pipeline when a pipeline fails. Projects that already have that automation are skipped, so a second run is safe. It also adds the GitLab MCP server to Claude Code, walks through browser sign-in, and verifies with a list_duo_agents_and_flows call. On Ultimate it asks whether to add security triage.
Prerequisites are specific. You need a GitLab.com top-level group with GitLab Credits, and Premium and Ultimate groups get promotional free credits each month per user. You need a project with a pipeline, open issues and at least one open merge request, plus Claude Code signed in, git, and the glab CLI authenticated with glab auth login. An Owner role configures the whole group; a Maintainer gets a project-only setup and a list of what to ask an Owner.
Four workflows from one terminal
The issue workflow: ask which issues are assigned to you, pick one, and tell the harness to assign it to the Developer flow. The session runs as a CI job on a runner, reads the issue and repository, writes the change and opens a merge request attributed to you rather than to a service account.
The review workflow: the Code Review flow reviews each merge request once it is marked ready. New groups and projects have it on by default; existing ones get it through the skill. A project setting under Settings > Merge requests > GitLab Duo Code Review can also start a new review on each push.
The pipeline workflow uses the automation described above. The security workflow, the fourth, needs Ultimate licensing.
Limits and open questions
GitLab names two blockers: runners that cannot run flows, and push rules enforced on the group. If the GitLab tools do not appear, check /mcp. A 403 at sign-in usually means MCP client access is off for the top-level group, which an Owner turns on with Allow connection to GitLab under Settings > General > Permissions and group features. Duplicate tools from user and project scope should be reduced to one.
The source does not say how many credits a flow consumes, what happens when credits run out, or what a review returns beyond findings on the merge request. The published post available to me ends mid-sentence in the code review section.
What it means for building with AI
Because flows run in GitLab, the same reviews, pipelines, scans and approval policies apply to agent output as to human work, and each MCP call generates an AI audit event. GitLab says the skill is harness-neutral by design because the repository's AGENTS.md requires it, so the setup survives a change of harness. That is the pitch: agents get speed, governance stays in place.
For teams that would rather start from a description than a repository, Draftship is an AI app builder where plain language becomes a built, hosted app with a database and user sign-in. Paid plans add source code access and GitHub push, so work can move into a repository later; the Draftship docs on GitHub and Vercel cover that path. Draftship is not affiliated with GitLab.
Primary sources
- Agentic engineering with GitLab and your local agent: A walkthrough · GitLab Blog, Nov 18, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.