Advanced GitLab MCP server

Advanced GitLab MCP server — 58 CQRS tools exposing 230 GitLab operations across 26 entity types. The tool catalog and parameters are filtered to each instance's GitLab version, tier, and token scopes, so the agent sees only what the connected instance actually supports.

Packages

This is a monorepo with two npm packages:

• @structured-world/gitlab-mcp — the MCP server. In-memory and file OAuth-session storage, no database dependency. This is all most users need.
• @structured-world/gitlab-mcp-db — optional PostgreSQL/Prisma OAuth-session backend for running multiple replicas behind a load balancer. Enable it with OAUTH_STORAGE_TYPE=postgresql; see the PostgreSQL deployment guide for the full setup.

Quick Start

{
  "mcpServers": {
    "gitlab": {
      "command": "npx",
      "args": ["-y", "@structured-world/gitlab-mcp"],
      "env": {
        "GITLAB_TOKEN": "your_gitlab_token",
        "GITLAB_API_URL": "https://gitlab.com"
      }
    }
  }
}

Requirements: Node.js >= 24

Highlights

• 58 tools across 26 entity types — projects, merge requests, pipelines, work items, wiki, and more
• CQRS architecture — browse_* for queries, manage_* for commands
• Connection resilience — Bounded startup, auto-reconnect with exponential backoff, disconnected mode when GitLab is unreachable
• Multi-instance support — Connect to GitLab.com, self-managed, and self-hosted instances with per-instance OAuth and rate limiting
• Multiple transports — stdio, SSE, StreamableHTTP
• OAuth 2.1 — Per-user authentication via Claude Custom Connector
• Read-only mode — Safe operation for production environments
• Auto-discovery — Detects GitLab config from git remotes
• Fine-grained control — Enable/disable tool groups, filter actions, customize descriptions
• Docker support — ghcr.io/structured-world/gitlab-mcp:latest

Tool model

GitLab exposes hundreds of API operations. Giving each its own MCP tool floods an agent's context; a thin API wrapper pushes that complexity back onto the agent. This server takes a third path: 58 CQRS tools, each holding several typed actions, expose 230 operations across 26 entity types.

• browse_* tools are read-only queries; manage_* tools are writes.
• Sub-resources fold into actions rather than new tools — browse_pipelines covers pipelines, jobs, and logs in a single tool.
• The tool list, actions, and parameters are resolved at schema-generation time against GitLab version, tier, token scopes, admin-mode, feature flags, and profiles.

Examples:

• browse_pipelines — list, get, jobs, job, logs, triggers
• manage_pipeline — create, retry, cancel
• browse_environments — list, get, list_deployments
• manage_environment — create, update, stop, delete, update_deployment_status
• browse_deploy_keys / manage_deploy_key — list, get / add, enable, update, delete

Documentation

Full documentation is available at gitlab-mcp.sw.foundation

Auto-generated Tool Reference

For the complete tool reference with parameters:

# View locally
yarn list-tools --detail

# Generate documentation
yarn list-tools --export --toc > docs/tools/api-reference.md

See the Full API Reference for the auto-generated tool documentation.

Docker

# HTTP mode
docker run -e PORT=3002 -e GITLAB_TOKEN=your_token -p 3333:3002 \
  ghcr.io/structured-world/gitlab-mcp:latest

# stdio mode
docker run -i --rm -e GITLAB_TOKEN=your_token \
  ghcr.io/structured-world/gitlab-mcp:latest

Connection Resilience

The server handles GitLab connectivity issues gracefully:

• Bounded startup — Server starts within GITLAB_INIT_TIMEOUT_MS (default 5s) regardless of GitLab availability
• Disconnected mode — When GitLab is unreachable (disconnected/failed state), only the manage_context tool is exposed, with local actions such as whoami, switch_profile, and set_scope for diagnostics. During active reconnect (connecting state), the full tool list remains available so MCP clients don't lose their tool catalog during brief outages. MCP clients are notified of tool availability changes via tools/list_changed
• Auto-reconnect — Exponential backoff reconnection (5s → 60s) with ±10% jitter
• Error classification — Transient errors (network, 5xx, timeouts) trigger auto-reconnect. Auth/config errors at startup transition to failed state (no auto-reconnect). Mid-session token revocation is detected via an authenticated HEAD /api/v4/user check that runs alongside each periodic health check (static token mode only; skipped in OAuth mode). A 401 or 403 on this check transitions the instance to failed state immediately.
• Instance health monitor — Each monitored instance URL has its own XState state machine. Untracked OAuth URLs currently pass through as reachable.

Feature Flags