Claude Code and Cursor on your repositories, with review that keeps up
Your developers use Claude Code, Cursor or Copilot, or soon will. They open more pull requests than the team can review, and that is where bugs and leaked keys get through. We spend two days with your team on your own repositories and leave behind rules, permissions and CI checks.
Smaller pull requests, and checks that catch what reviewers miss
An agent can open more pull requests in a morning than your team can properly review in a day. If review stays the way it is, PRs queue up, or people start approving code they haven't read. So we work with smaller PRs, a rules file that prevents the common mistakes, and CI that catches what a tired reviewer misses.
Two days, on your code
On site in Ghent, or as four remote half-days.
Day one: set it up
- How you work now. Your repo layout, branching, CI, and how code review happens today.
- The rules file. Written together: CLAUDE.md, AGENTS.md or
.cursor/rules/, depending on what you use. - Permissions. Which commands the agent runs without asking, which ones need a person, which files it never reads.
- Secrets and CI. Where credentials live, what the agent can reach, and the checks every PR has to pass.
Day two: use it
- Pairing on your backlog. Real tickets, picked by you. Each developer drives the agent; we sit next to them.
- Reviewing what came out. We review those PRs together and turn what we find into rules and checks.
- The team agreement. When an agent may push, who merges, what a PR description says. Written down in the repo.
A rules file written for this repository
The agent reads it at the start of every session, so it only says what is specific to your code.
## StackTypeScript, Fastify, Postgres 16. Run `pnpm test` and `pnpm lint`before you say a task is done.## Rules- Every query on `invoices` filters on `tenant_id`. No exceptions.- Schema changes go in a new file in `migrations/`.Never edit a migration that is already merged.- Ask before adding a dependency. Say why the code we have is not enough.- Never delete, skip or loosen a test to make it pass.- Never read `.env*` files. Ask for the variable name instead.- Work on a branch. Open a draft PR. A person merges.## In the PR description- What you changed, and which files you did not expect to touch.
{"permissions": {"allow": ["Bash(pnpm test *)", "Bash(pnpm lint)"],"ask": ["Bash(git push *)"],"deny": ["Read(**/.env*)"]}}
What to look for in a PR an agent wrote
- New dependencies. Check that the package exists and is the one you meant. Assistants sometimes suggest names that don't exist, and an attacker can register them.
- Tests that changed to pass. A loosened assertion, a skipped case, a mock that replaces the thing under test.
- Access checks. A new endpoint without the auth middleware, a query without the tenant filter.
- Files outside the ticket. Changes nobody asked for, mixed into a PR about something else.
The person who prompted the agent is the author. They read every line before asking for review, and the PR description says what they checked.
Keep secrets out of the agent's reach
An agent runs commands in your shell, with your credentials.
- No production credentials on developer machines. Development keys with the least access that works.
.envfiles denied in the permissions. That covers the files the agent reads itself; agrep -ror a script can still get at them. Real secrets belong in a secret manager.- Secret scanning before the commit and again in CI, with gitleaks or your Git host's push protection.
- MCP servers reviewed like dependencies. Each one is code that runs with your permissions.
When an agent may push
Our default. We adjust it with you, then enforce it in the repository settings.
- Work on its own feature branch, run the tests, push that branch
- Open a draft pull request
- Push to
main, or merge anything - Deploy, or run migrations against a shared database
- Change CI configuration or branch protection
CI backs this up: lint, type check and tests required before merge, secret scanning on every push, CODEOWNERS on auth, payments and migrations, and at least one human approval.
What you pay us, and the licences in your own name
We work at a fixed day rate or at a fixed price per outcome. You know the price before we start. You buy the Claude Code or Cursor licences directly from the provider.
- It works best for a team that already has a repository, CI and code review. If there is no CI yet, part of day one goes to that.
- Up to 4 developers per programme
When this isn't the right fit
- You built an app with Lovable, Bolt or Cursor and aren't a developer yourself: a vibe-coded app check fits better.
- Your team doesn't review code today. An assistant makes that worse. Agree on code review first.