Rolling Out Claude Code in the Enterprise: Security, Settings and Method
How to roll out Claude Code across a mid-size or large company: managed settings, permissions, data, pilot and measurement. A practical guide backed by official docs.
Installing Claude Code on one laptop is quick. Rolling it out to 50 or 500 developers, under rules your security team accepts, is a different project.
Here's what to decide before rollout: where settings live, what the agent can do, which data it sees, and how to tell if it's worth it. Everything about the product comes from the official Claude Code documentation. It changes fast, so check it again when you deploy.
Four decisions before you hand out access
- Plan and authentication. Server-managed settings are limited to Claude for Teams and Claude for Enterprise. With individual API keys, you lose central control.
- Who owns the rules. On Anthropic's side, only Owner or Primary Owner roles can edit managed settings. Decide who holds them internally (usually platform or security, not each team).
- What the agent runs without asking. This is the most sensitive call. Shell commands, network access, off-limits files.
- How you measure. If you don't set metrics upfront, the pilot ends on a gut feeling.
Managed settings: server or endpoint?
Claude Code reads settings from several files, from personal to global, and an organization can enforce its own on top (Settings docs). There are two ways to enforce them:
| Approach | Best for | How it works |
|---|---|---|
| Server-managed settings | Companies without MDM, unmanaged devices | Configured in the claude.ai console, fetched at startup and refreshed hourly |
| Endpoint-managed settings | Companies with MDM or endpoint management | Deployed through OS policies (macOS managed preferences, Windows registry) or managed settings files |
Source: Server-managed settings. Devices need to reach api.anthropic.com for the server option.
Our advice: if you already run MDM, use it. Your rules then follow the same path as the rest of your security policy.
Permissions: allow, ask, deny
Claude Code separates what it can do alone (allow), what needs confirmation (ask) and what is blocked (deny) (Permissions docs). A sensible baseline:
| Category | Starting rule | Why |
|---|---|---|
| Reading repository code | allow | That's the core use |
| Tests, lint, build | allow | Known commands, no external effect |
| Writing files | ask at first, allow later | Teams build trust during the pilot |
.env files, secrets, keys | deny | The agent doesn't need them to code |
git push, deploys, production commands | deny or ask | Keep a human on anything leaving the laptop |
| Network to unknown domains | ask | Limit leaks and surprise dependencies |
Put these rules in managed settings. Developers can tighten them, not loosen them.
Data: what the agent actually sees
The agent sends the model the context it needs: code snippets, command output, files it reads. Before the pilot, ask your DPO and CISO three questions:
- Which repositories are in scope? Start with ones holding no personal data and no plaintext secrets.
- Which retention terms apply? They depend on your contract. Read them in your Anthropic commercial terms, not in a blog post (this one included).
- What about sensitive data? For health data in France, HDS rules apply on top. See our article on AI and health data.
The pilot: one or two teams
Don't switch everyone on the same day. A pilot that holds up:
- Pick one or two volunteer teams, on representative but non-critical repos.
- Write a
CLAUDE.mdper repo: conventions, test commands, known pitfalls. Claude Code reads it at startup (Memory docs). It's often what separates a useful agent from one that improvises. - Set two or three metrics before starting: ticket-to-PR time, share of PRs passing review without major rework, time spent on repetitive tasks (migrations, tests).
- Review weekly what got in the way: permissions too strict, missing context, poorly chosen tasks.
- Decide at the end: expand, adjust, or stop.
Claude Code, Codex, or both?
Many companies also trial OpenAI's Codex, which has its own sandbox and approval modes. The choice mostly depends on your existing contracts and data policy. The method stays the same: central settings, explicit permissions, a measured pilot. If teams use several agents, keep equivalent rules everywhere. Otherwise usage drifts toward the least governed tool.
What we do at The Vibe Company
The Vibe Company is an AI agency in Paris, founded in 2024 by Stan Girard and Antoine Dewez (ex-Theodo, ex-Stockly), backed by Y Combinator (W24). We've run dozens of projects, with three specialties: HDS and security, frontend, DevOps.
We use these coding agents every day to ship. On a rollout, we set up managed settings and permissions, write CLAUDE.md files with your teams and train people during the pilot. Useful for procurement: we don't claim any security certification.
FAQ
Do I need an Enterprise plan to roll out Claude Code?
Not to use it, but server-managed settings require Claude for Teams or Enterprise. Without them, use your MDM or managed settings files.
Can a developer bypass the rules?
Managed settings take precedence over user and project settings, per the documentation. That doesn't replace the usual controls: repo access rights, branch protection, code review.
How long does a rollout take?
Setup first, then a pilot before you expand. How long it takes mostly depends on your security sign-off process.
Let's talk about your rollout
Want to scope a Claude Code or Codex pilot for your teams? Book 30 minutes with Stan.
