Use GitLab integration to query and manage projects, merge requests, issues, pipelines, branches, commits, files, and releases on the configured GitLab instance. Use whenever the user asks about GitLab, listing MRs, reviewing diffs, opening or merging MRs, creating/closing issues, comparing branches, reading pipeline status, fetching files, searching code, checking "my open MRs" or "my assigned issues", etc. Use even when you think you know the answer, repository state is dynamic and only the live API reflects current MR status, pipeline runs, or issue assignments. Do not use for plain git operations on the local checkout, GitHub.com (use github), general code review questions, or anything outside the configured GitLab instance.
Instalação
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Use GitLab integration to query and manage projects, merge requests, issues, pipelines, branches, commits, files, and releases on the configured GitLab instance. Use whenever the user asks about GitLab, listing MRs, reviewing diffs, opening or merging MRs, creating/closing issues, comparing branches, reading pipeline status, fetching files, searching code, checking "my open MRs" or "my assigned issues", etc. Use even when you think you know the answer, repository state is dynamic and only the live API reflects current MR status, pipeline runs, or issue assignments. Do not use for plain git operations on the local checkout, GitHub.com (use github), general code review questions, or anything outside the configured GitLab instance.
GitLab access goes through MCP Hub. Claude does not see gitlab__* tools directly: use search_tools to discover them and execute_code with the injected gitlab global to invoke. host_url and token are pre-configured at the worker; never pass them, never ask the user for a GitLab token.
Workflow
search_tools({ query: "<keywords>", detail_level: "names_only", service: "gitlab" }): discover what is available.
search_tools({ query: "<toolName>", detail_level: "full_schema", service: "gitlab" }): get the exact parameter schema.
execute_code: call the tool using the parameter names from the schema.
Always run step 1-2 before guessing a tool name. The live schema is the source of truth.
Pitfalls
IID vs DB ID. Issues and MRs use per-project issue_iid / mr_iid, the numbers users see in URLs (/-/issues/42, /-/merge_requests/42). Not the internal database ID. Accepts a plain number or a string, with or without a leading # (e.g. 42 or "#42").
Project paths. Pass group/subgroup/project as a plain string. The tool URL-encodes slashes for you; do not pre-encode %2F.
Identity ("my MRs"/"my issues"/"my projects").listMrIds/listIssues' scope: "assigned_to_me"/"created_by_me" and listProjectIds's owned: true resolve identity server-side without needing a username, but scope filters by assignee, not reviewer: "MRs I need to review" needs reviewer_username plus the current-user tool instead. The tool description names that current-user tool when you need the caller's own numeric ID or username directly (e.g. to compare against an author_id field returned by another tool).
Pipeline triggers cost CI minutes. Confirm before retrying or triggering a pipeline.
Search beats tree-walking. When looking for a symbol or file, reach for searchCode before listing directories blindly.
Job IDs come from pipelines.getJobLog, downloadArtifact, and deleteArtifacts take a job_id, get valid values from getPipelineFull's jobs array, not by guessing. deleteArtifacts is irreversible and erases the job log too, not only the artifacts: treat it with the same confirmation weight as any other destructive delete.
downloadArtifact returns a job log, not the CI artifact zip. This worker cannot fetch raw artifact bundle contents.
Write/delete confirmation. Per claude-resources/CLAUDE.md, every create/update/close/merge/delete needs explicit user approval before execution. Read operations (list/get/search/compare/diff/blame) never need confirmation.
When NOT to use
Plain git operations on the local checkout (clone, commit, push): use the shell, not this integration.
GitHub.com repositories: use the github integration instead.
Anything outside the single GitLab instance configured for this project.
