SystemFlare
ES Talk to an engineer

AI coding agent governance

Roll out AI coding agents without losing control

Your developers already use AI agents. We set the rules so the agents can be fast without being able to leak secrets, touch production or merge unreviewed code.

AI coding agent governance is the engineering setup that lets a team use Claude Code, Codex, Cursor and similar agents safely: central permission policies, managed settings, agents that can only open pull requests, pre-merge security checks, secret isolation, an approved list of MCP servers and an audit trail. It is for engineering leaders rolling out agents across a team, especially in regulated companies. It starts at €4,000 and takes 2–3 weeks.

Who it is for

  • CTOs and engineering managers rolling out Claude Code, Codex or Cursor to a team
  • Regulated companies (finance, payments, ICT suppliers) that must explain to auditors how AI touches their code
  • Security teams asked to approve AI agents without a clear way to control them
  • Teams that already had an agent do something unexpected

Why now

  • Developers are using AI agents on their own laptops with their own settings, and nobody knows what they can access
  • An agent ran a command, installed a package or connected an MCP server that nobody approved
  • Your DORA or PCI DSS evidence says every production change is reviewed, and an agent with a deploy key would break that
  • The EU AI Act transparency duties apply, and management wants a written position on how AI is used

What you get

  • Agent policy: what agents may read, run and change, written for your repositories and reviewed with your team
  • Managed settings deployed centrally: allowed and denied commands, file paths and network access, which developers cannot weaken locally
  • PR-only agents: agents work on branches and open pull requests; they never push to protected branches or deploy
  • Pre-merge checks: secret scanning, forbidden files, dependency checks for invented or typosquatted packages, and flags for risky changes
  • Secret isolation: agents get no production credentials, no server access and no database access
  • Approved MCP servers and tools: an allowlist, with the reasons for each, and a way to request new ones
  • Repository rule files (such as CLAUDE.md or AGENTS.md) that tell agents your conventions and limits
  • Audit trail: which agent proposed what, who approved it, and what was deployed, kept in Git and CI history

How it works

  1. Inventory

    Which agents, tools, MCP servers and credentials are in use today, and on which repositories.

  2. Policy

    A short, practical policy agreed with engineering and security: what agents may do, and what always needs a human.

  3. Enforce

    Managed settings, branch protection, pre-merge checks and secret isolation, rolled out through pull requests.

  4. Train and hand over

    A session with the team, rule files in each repository, and a review of the first weeks of agent work.

The principle: rules live in the pipeline, not in the prompt

A well-written prompt or rule file makes an agent behave better. It does not make it safe. Agents misunderstand, follow instructions hidden in files they read, or simply make mistakes. So we put the important limits where an agent cannot change them:

  • Branch protection decides what can merge, not the agent.
  • Checks in CI decide whether a change is acceptable: syntax for the target runtime, secrets, forbidden files, dependency checks and a flag on risky changes such as config, CI, dependencies, SQL or data-deleting code.
  • The deploy gate runs on the server and cannot be weakened by a pull request.
  • Credentials for production simply do not exist in the agent’s environment.

How we run it ourselves

We use AI agents in our own delivery, and the setup we build for clients is the one we use. The agent works only through pull requests. It has no server, database or secret access. It cannot merge with red checks. Every request, change and approval stays in Git history. This is how a shop owner in one of our projects requests changes from Claude while the production gate stays in force.

What we do not do

We do not pick your AI vendor or sell licences, and we do not ban tools. We make the tools you choose safe to use, and leave your team with rules they understand.

Related work

Questions clients ask

How do we roll out Claude Code or Cursor to our team safely?

Start with three rules enforced by tooling, not by trust: agents only open pull requests, agents never hold production secrets, and nothing merges without green checks and a human review. Then add managed settings and an approved list of tools. That is what this service sets up.

How much does AI agent governance cost?

From €4,000 excluding VAT for one team and its main repositories, usually in 2–3 weeks. You get a fixed quote after a first call.

Will this slow our developers down?

Very little. Agents can still read code, run tests and propose changes freely. The limits sit at the points that matter: secrets, production and merging. Most teams find reviews easier, because every agent change arrives as a clean pull request.

Do you use AI agents yourselves?

Yes. We use Claude in our own work. The agent works only through pull requests, never has server, database or secret access, and never merges with red checks. We set up the same rules for clients.

What are MCP servers and why do they need approval?

MCP servers connect an agent to other systems: databases, ticketing, cloud accounts, browsers. Each one widens what the agent can see and do. An allowlist makes sure every connection is a decision, not an accident.

Does this help with auditors?

Yes. Auditors ask how changes are controlled. With PR-only agents, branch protection and checks, the answer is the same for humans and AI: every change is reviewed, tested and recorded in Git.

Using AI agents already? Let's make it controlled

Tell us which tools your team uses and how many repositories are involved. We will propose a policy and a fixed price.

Talk to an engineer