How do you roll out Claude Code to an engineering team safely?

Roll out Claude Code by choosing the provider first, deploying managed settings before developers get access, piloting on a few repositories with permission rules, sandboxing and an MCP allowlist in place, then expanding while analytics, spend limits and OpenTelemetry track usage. Managed settings override anything a developer configures, so your policy belongs there, not in a wiki page people may not read.

Claude Code is the agentic coding tool in Claude, powered by Anthropic. It reads a codebase, edits files, runs commands and opens pull requests, in the terminal, VS Code and JetBrains, the desktop app and the browser. That power is why it needs its own rollout track, separate from chat. This guide follows Anthropic's administrator documentation as of October 2026 and adds the order we recommend. Setting names are exact so your platform team can act on them.

Which way should developers access Claude Code?

The provider decides billing, authentication, which compliance posture you inherit, and which features exist. Anthropic's default recommendation is a Claude Team or Enterprise plan, which puts Claude Code and the Claude apps under one subscription with no infrastructure to run.

What is the rollout sequence?

What should managed settings enforce on day one?

Managed settings sit at the top of the precedence order: managed, then command line, then project local, shared project and user settings. Nothing a developer sets overrides a managed key. List keys such as permissions.allow and permissions.deny merge across files, so developers can add to your lists but not remove from them. These are the controls Anthropic documents, with our suggested starting position.

How does policy reach developer machines?

Claude Code checks four managed sources in priority order. Server-managed settings come from the claude.ai admin console, require a Team or Enterprise plan, are fetched at startup and refreshed hourly, and need no endpoint tooling. Next is the macOS plist (com.anthropic.claudecode) or the Windows HKLM registry key (HKLM\SOFTWARE\Policies\ClaudeCode), both of which need admin rights to change. Then a managed-settings.json file: /Library/Application Support/ClaudeCode/ on macOS, /etc/claude-code/ on Linux and WSL, C:\Program Files\ClaudeCode\ on Windows. Last is the Windows HKCU key, which users can write without elevation, so Anthropic says to treat it as a convenience default, not enforcement. If you mix providers, use server-managed settings for claude.ai users plus a file or MDM fallback for everyone else. Cloud sessions on the web receive only server-managed settings.

Should you allow auto mode?

Auto mode is now the built-in starting permission mode for interactive terminal and VS Code sessions. A separate classifier model reviews actions and blocks those it judges unsafe, while your ask and deny rules still apply. Manual mode starts read-only and asks before editing files or running commands, apart from a built-in set of read-only commands such as ls, cat and git status. Auto mode reduces prompt fatigue, which matters because people who click approve all day stop reading. Manual mode keeps a human on every action. A reasonable pattern is to allow auto mode with sandboxing on and strong deny rules, and to require Manual mode in repositories that touch production infrastructure or regulated data. Whatever you choose, set it in managed settings rather than leaving it to each developer.

How should each repository be set up?

• Commit .claude/settings.json with the team's permissions, hooks and plugins. Allow rules and most env values in it apply only after each developer trusts the folder; deny and ask rules apply immediately.
• Keep personal overrides in .claude/settings.local.json, which Claude Code keeps out of commits when it creates the file. If someone creates it by hand, add it to .gitignore.
• Write a CLAUDE.md with build and test commands, architecture notes and conventions. Anthropic suggests keeping it under 200 lines and moving workflow-specific instructions into skills, which load only when needed.
• Review .mcp.json like code. It is not the only source of MCP servers: user-scope servers, claude.ai connectors and plugins can add more, which is why the managed allowlist matters.
• Put repeatable team workflows, such as a pull request checklist or a release-notes format, in .claude/skills or a company plugin.

How do you run the pilot and measure it?

Choose two or three teams and repositories with good test coverage, since Claude Code works best when it can check its own work. Start with bounded task types: writing tests for untested code, fixing lint errors, dependency updates, small bug fixes, and first-pass code review. On Team and Enterprise, the analytics dashboard at claude.ai/analytics/claude-code shows daily active users and sessions, lines of code accepted and suggestion accept rate. Connect the Claude GitHub app and turn on GitHub analytics to add contribution metrics: merged pull requests with Claude Code assistance, lines shipped, and a leaderboard. Anthropic describes these metrics as deliberately conservative. They cover github.com repositories only, and are not available with zero data retention enabled. Compare them with your existing measures, such as DORA metrics or cycle time, rather than reading them alone.

What will it cost, and how do you control spend?

Anthropic's cost documentation says that across enterprise deployments the average is around $13 per developer per active day and $150 to $250 per developer per month, with 90% of users below $30 per active day. Your numbers depend on model choice, codebase size and automation, so establish a baseline in the pilot. On Team and Enterprise seat plans, each member's usage draws from a seat allowance that resets on a five-hour window and a weekly window and is shared with chat. To let people keep working past it, turn on usage credits and set spend limits at organization, group or member level in Organization settings > Usage. On Console, set workspace spend limits and rate limits; Anthropic suggests about 15,000 to 20,000 tokens per minute per user for 100 to 500 users and 10,000 to 15,000 above 500. On cloud providers, use your cloud budget controls plus OpenTelemetry or a gateway for per-user attribution.

• Default developers to Sonnet and reserve Opus for complex design and multi-step reasoning; Anthropic notes Sonnet handles most coding tasks for less.
• Teach /clear between unrelated tasks, plan mode for complex changes, and specific prompts. Long sessions resend the whole conversation with every request.
• Watch agent teams: Anthropic estimates roughly 7x the tokens of a standard session when teammates run in plan mode. They are off by default.
• Export OpenTelemetry metrics to your observability stack; it is the only option that streams per-user token and cost data in near real time on every provider.

What about code and data handling?

On Team, Enterprise, Claude API and cloud provider plans, Anthropic does not train models on your code or prompts. Zero data retention is available to qualified Claude Enterprise accounts. For HIPAA-ready Enterprise organizations, the Primary Owner must apply the HIPAA configuration to Claude Code in local mode for the business associate agreement to cover it, the VS Code and JetBrains extensions are not covered, and sessions routed through a gateway are not eligible. Cloud sessions run in isolated Anthropic-managed VMs, keep GitHub credentials out of the VM, and reject branch deletions and tag pushes. If you need request-level audit logs, Anthropic recommends a gateway between developers and the provider.

Where do Claude Code rollouts go wrong?

• Handing out seats before policy is deployed, then trying to claw back permissions developers already rely on.
• Relying on deny rules alone without sandboxing, which leaves curl and wget open.
• Using the Windows HKCU registry key as the enforcement channel; users can change it.
• Leaving forceLoginMethod unset, so developers use personal subscriptions or old Console keys outside your controls.
• Allowing any MCP server or plugin marketplace. Anthropic reviews directory connectors against listing criteria but does not security-audit or manage any MCP server.
• Measuring success by lines of code alone. Track review load, defect rates and cycle time as well.
• Skipping training. Free Claude Academy courses (Claude Code 101 and Claude Code in Action) plus a team session on your own repo work better than a link in Slack.

Abstrakt Solutions sets up Claude Code for engineering organizations, including managed settings, sandbox policy, plugins and team training, and uses it ourselves for Salesforce development. See our Claude Code for engineering teams services.