ReviewGate runs inside your own network, and the routine of installing it can be handed to an AI assistant — Claude Code, Cursor, a local model, any of them. You give it a link to the instructions and answer a couple of questions; it brings the containers up, sets up the webhook and checks that everything works.
the playbook is written for GitLabThere is no separate step-by-step playbook for GitHub yet. If you host on GitHub, the agent takes the token and webhook steps from Connecting GitHub; the rest (the image, .env, the run, the check) is the same.
why botherThe longest stretch of onboarding is the manual fiddling with docker, .env, the webhook and config.yml. An agent does it in minutes and runs the diagnostics itself. The instructions are plain text that cloud and local models read equally well.
How to start
Pick the route that fits your tool:
An agent with internet access (Claude Code, Cursor). Give it the link https://reviewgate.dev/llms.txt — it fetches the playbook itself and walks the steps.
A local or offline model. Download the playbook ai-setup.md and hand the file to the agent locally — it needs no internet beyond access to your GitLab and your model.
A chat assistant. Copy the preamble below into the chat — it sets the boundaries and points at the playbook.
preamble for an AI assistant
Set up ReviewGate (AI code review for our self-hosted GitLab)
following these instructions: https://reviewgate.dev/ai-setup.md
Rules:
— do not invent secrets (the GitLab token, the model key) — ask me for them;
— do not send our code or diffs anywhere, work only inside our network;
— show me the commands before running them.
What the AI will ask you for
You type the secrets yourself — the agent neither invents them nor sends them anywhere.
what
kind
the URL of your GitLab and the project
a decision
the bot token (api scope)
a secret — you type it
a cloud model or a local one
a decision
the model key and the proxy, if you use a cloud model
a secret — you type it
how many reviews to run in parallel (REVIEW_CONCURRENCY) — by team size and provider limits
a decision
What the AI does on its own
downloads the client docker-compose.yml and writes an .env from your answers;
brings the containers up and checks /api/health;
creates the webhook in GitLab through the API — the step people most often forget by hand (the webhook secret is generated automatically and written to both sides at once);
proposes a .reviewgate/config.yml with team rules derived from your linters and your ADRs;
runs diagnose.sh and checks a review on a test merge request — including the «🔬 Run diagnostics» block in the summary (switched on for the duration of the install): it shows that the config was picked up and that both model passes ran;
optionally installs the review for itself: the same engine is available to the agent locally — the CLI, the pre-git push hook and MCP — so it can check code against your team's rules before a merge request even exists.
Privacy guarantees
the code stays with youThe agent works inside your network and, by the instructions, sends out neither sources, nor the diff, nor the contents of .env. You type the secrets. The playbook is open — you see exactly what the AI will read: ai-setup.md, llms.txt.