A practical guide to the new admin controls for Claude Code in 2025

Let’s be honest, it’s a tricky balancing act. On one hand, you have powerful AI coding tools like Claude Code that can give your development teams a serious boost. On the other, you have the absolute necessity of keeping your systems secure, compliant, and on budget. Letting developers run loose with an AI that can access your codebase and execute commands can feel a lot like handing over the keys to the entire building.

Thankfully, Anthropic has rolled out a set of enterprise features and admin controls to tackle this very problem. This guide will walk you through what these controls are, why they matter so much, and how you can use them to bring Claude Code into your organization without the accompanying stress.

What is Claude Code?

Before we get into the nitty-gritty of the controls, let’s quickly get on the same page about what Claude Code is. It’s an AI coding assistant from Anthropic that lives right in a developer’s terminal. You can think of it as a pair programmer you can chat with in plain English.

It’s there to help with analyzing codebases, writing new code, planning projects, and handling repetitive tasks. The main thing to grasp is that because it operates in the terminal, it has the same permissions as the developer using it. It can read files, write code, and run commands just like a person would. That capability is what makes it so useful, but it’s also precisely why having solid admin controls for Claude Code is a must for any business.

A screenshot of Claude Code operating within a developer's terminal, demonstrating its native integration.

Why robust admin controls are non-negotiable

While Claude Code can be a huge help for productivity, that power comes with some real risks if it isn’t managed correctly. You can’t just switch it on and hope for the best.

Security risks without admin controls

An AI assistant with direct access to your systems can cause problems, even if its intentions are good. Here are a few scenarios to think about:

• Accidental (or malicious) commands: A slightly off or poorly phrased prompt could trick the AI into running a destructive command. Imagine it misunderstanding a request and trying to run "rm -rf /" or using "curl" to send sensitive data somewhere it shouldn’t.

• Leaking sensitive data: Without any guardrails, Claude Code can read any file the user has access to. This includes sensitive files like ".env" configurations, SSH keys, or AWS credentials. A simple prompt asking to summarize a project could accidentally pull in and display these secrets.

• Unintended system changes: The AI might suggest installing an unvetted dependency that contains malware. Or it could try to modify critical configuration files or interact with production systems without the proper approvals in place.

An illustration of the security guardrails that provide admin controls for Claude Code, preventing unauthorized actions.

The compliance and governance challenge

For companies in regulated industries like finance or healthcare, the stakes are even higher. You need a clear record of what your AI is doing, which files it’s accessing, and what code it’s generating to stay compliant. Without audit trails, you’re essentially flying blind.

And then there’s the cost. Uncontrolled AI usage can lead to some eye-watering bills at the end of the month. Admins need a way to see what’s being used and manage spending, which has been a hot topic since Anthropic first introduced rate limits.

A deep dive into Claude Code’s admin controls

Anthropic has built a layered system to give organizations the control they need. It starts with high-level account management and goes all the way down to command-by-command permissions.

High-level organizational controls for enterprise and team plans

The first layer of control is at the plan level. With the Team and Enterprise plans, admins get a central dashboard to manage everything.

Here’s what you get:

• Self-serve seat management: Admins can easily add, assign, and manage user seats. You can give out standard seats for the Claude.ai web app and premium seats for developers who need Claude Code.

• Granular spend controls: You can set spending limits for the entire organization and for individual users. This is a big deal for keeping costs predictable.

• Usage analytics: The dashboard shows you how your team is using the tool. You can track metrics like lines of code accepted and suggestion acceptance rates, which helps you figure out if you’re getting your money’s worth.

• SSO and Role-Based Access: Standard enterprise features like Single Sign-On (SSO) and role-based permissions help you centralize access and make sure only the right people have admin privileges.

Granular terminal-level controls with settings.json

This is where you can get really specific. Using simple JSON configuration files, you can control exactly what Claude Code is allowed to do on a developer’s machine. The system works with a clear hierarchy:

1. Enterprise Managed Policies ("managed-settings.json"): This is the master file, typically handled by IT or DevOps. These settings are enforced across the company and can’t be changed by individual users. It’s your main safety net.

2. Project Settings (".claude/settings.json"): These settings live within a project’s repository and are shared by the team, making it a good spot for project-specific rules.

3. User Settings ("~/.claude/settings.json"): This is for a developer’s personal preferences, as long as they don’t clash with the project or enterprise rules.

The system is built around a set of permission rules. Here’s a quick look at how they work:

Permission Type	Description	Example
"allow"	A whitelist of commands or actions that can run without asking. Best for safe, everyday tasks.	["Bash(npm run lint)"]
"ask"	Actions that need the user to say "yes" every single time. Good for anything potentially risky.	["Bash(git push:*)"]
"deny"	A blacklist of actions that are completely forbidden. This is the strongest rule and can’t be bypassed.	["WebFetch", "Read(./.env)"]

Using these rules, you can stop Claude Code from reading sensitive files, running dangerous commands, or accessing the internet. It gives you very precise control.

A code snippet showing the "settings.json" file used for granular admin controls for Claude Code.

The Compliance API for advanced observability

For organizations with strict compliance requirements, Anthropic offers a Compliance API on its Enterprise plan. This is more than just a dashboard; it gives your compliance teams real-time, automated access to usage data and conversation logs.

This means you can build your own continuous monitoring systems, feed Claude’s activity logs into your existing security tools, and automate policy enforcement. It’s a serious feature for businesses that need a complete audit trail of all AI activity.

Claude Code pricing and plans

Access to Claude Code is part of the premium seats on the Team and Enterprise plans. The individual plans you see on the website (Free, Pro, Max) are just for the web app, not for this kind of setup.

• Team Plan: Starts at $20 per user/month (billed annually) for standard seats. To get premium seats with Claude Code, you’ll need to talk to their sales team.

• Enterprise Plan: This involves custom pricing. You have to contact their sales team to get a quote based on what your organization needs. This plan includes all the advanced features like the Compliance API, SSO, and higher usage limits.

It’s worth noting that the pricing for premium seats and any extra usage isn’t public, so a conversation with their sales team is necessary to figure out the final cost.

A practical guide to implementing Claude Code controls

Knowing about the controls is one thing, but using them effectively is another. Here’s a practical way to get started.

Best practices for configuring your settings

Here’s a sensible approach to locking down your configuration, based on common security practices:

• Start by denying everything. Before you allow anything, explicitly block access to sensitive files and network commands. Think ".env" files, "secrets/" directories, and tools like "curl" or "wget".

• Make "ask" your default. For any command that modifies files, runs builds, or pushes code, set it to "ask" for permission every time. This simple step prevents the AI from doing something unexpected.

• Be careful with what you "allow"list. Only add commands to your allowlist if they are 100% safe and used frequently, like running a linter or a specific, sandboxed test script.

• Review your settings regularly. Don’t just set it up and walk away. Check in on your "managed-settings.json" and project-level settings every so often to make sure they’re still appropriate.

This video offers practical tips on setting up and using Claude Code, which is helpful when implementing admin controls for Claude Code.

Limitations: What the controls don’t solve

These controls are a great step, but they don’t solve every problem. It’s worth remembering that some users have pointed out that the usage rate limits still apply even to premium seats, which can be frustrating for developers on a tight deadline.

There’s also the complexity to consider. Setting up and maintaining these JSON configurations isn’t trivial; it requires technical know-how. It’s a system designed for developers, which can create some management overhead for the IT and DevOps teams tasked with creating and enforcing these policies.

Beyond coding: Simplifying AI controls for your entire organization

The manual, code-based setup for Claude Code is powerful for development teams, but what about the rest of your business? Your customer support, IT, and HR teams also need AI tools, but you can’t expect them to start managing JSON files. This is where the need for a different kind of control panel comes into play.

This is exactly the problem we set out to solve with eesel AI. Our thinking is that powerful AI should also be simple and safe for anyone to manage.

• Genuinely Self-Serve: Forget about complicated setups or mandatory sales demos. With eesel AI, you can connect your help desk like Zendesk or Intercom, hook up knowledge sources like Confluence or Google Docs, and have an AI agent running in minutes, all from a clean, straightforward dashboard.

• Total, User-Friendly Control: We give you fine-grained control without asking you to write a single line of code. A support manager can easily decide which topics the AI should handle, which knowledge sources it can use, and what actions it can take (like escalating a ticket or looking up an order in Shopify), all with a few clicks.

• Risk-Free Simulation: One of our most useful features is the ability to simulate how your AI agent will perform on thousands of your past support tickets. You can check its responses, see its resolution rate, and prove its value before it ever interacts with a live customer. This takes all the guesswork out of the process and lets you deploy AI with complete confidence.

eesel AI is built for teams that need the power of a custom-trained AI but want the simplicity and safety of an enterprise-grade, no-code platform.

Balancing power and responsibility with Claude Code’s admin controls

The new admin controls for Claude Code are a significant step forward for using agentic AI safely in a business setting. They provide a solid framework for letting developers be productive while maintaining security and oversight. A successful rollout, however, really depends on a thoughtful, security-first approach that combines high-level policies with detailed, in-terminal permissions.

Ultimately, whether you’re automating coding tasks or customer support, the future of enterprise AI belongs to platforms that deliver both incredible power and the intuitive, robust controls needed to manage it all responsibly.

Ready to deploy AI with controls that anyone can manage? Try eesel AI for free and see how easy enterprise-grade AI can be.