Integration guide
GitLab Integration
DevSecOps platform where AI features and integrations can connect model assistance to repositories, issues, and delivery workflows.
Recommended path
GitLab local AI quick start
- 1. Confirm the GitLab licence and deployment boundaryVerify that the required GitLab and Duo add-ons are available, then decide whether the deployment must be fully offline or can keep selected GitLab-managed services.
- 2. Deploy the AI Gateway and model endpointRun the self-hosted AI Gateway with a reachable OpenAI-compatible /v1 server. GitLab documents vLLM 0.18.1 or later for offline production deployments.
- 3. Add, assign, and test each feature modelRegister the endpoint, assign models to the intended Duo features, and test code completion and chat separately before enabling the configuration for a wider group.
Best for
Teams using GitLab for source control, CI/CD, planning, and software delivery.
Model support
GitLab Duo Self-Hosted serves Duo features from your own infrastructure through a self-hosted AI Gateway. It accepts any model exposed through an OpenAI-compatible /v1 endpoint — GitLab names vLLM as its documented serving platform, and separately publishes an end-to-end solutions guide for deploying Duo Self-Hosted against Ollama.
Choose models by role
Code completion
GitLab-compatible completion model
Use a model and serving configuration listed for the code-completion feature. A general chat model is not automatically a valid completion replacement.
Chat and agent features
GitLab-compatible instruction model
Choose an instruction model supported for the specific Duo feature, with tool use where that feature requires it. GitLab's compatibility matrix is feature-specific.
How to use this integration
- Serve Duo features from your own vLLM deployment so code never leaves your infrastructure
- Meet data-residency or air-gap requirements that rule out a vendor-hosted assistant
- Standardise on one self-hosted model across merge requests, issues and the IDE
- Evaluate whether the Duo Enterprise add-on cost is justified before committing to a rollout
Connecting a local model
Deploy the self-hosted AI Gateway and point it at your own OpenAI-compatible endpoint; GitLab documents vLLM (0.18.1 or later) for this path, and its solutions guide walks through an Ollama deployment on AWS or Google Cloud. This is infrastructure work rather than a settings change.
Tradeoffs
Read the licensing per surface rather than as one rule: Duo Self-Hosted needs Premium or Ultimate plus the seat-based Duo Enterprise add-on, while Agent Platform Self-Hosted needs its own add-on with Enterprise License Agreement billing on offline licences and usage-based credits on online ones. The support status also differs from the deployment story — the Ollama walkthrough is a solutions guide positioned for testing and evaluation, and GitLab points at more powerful GPU instances for production, so treat it as a documented path rather than a supported production configuration. Hybrid setups keeping some GitLab-managed models still require internet connectivity, which GitLab states is not fully self-hosted or isolated.