Anthropic is giving enterprise developers a more controlled way to use Claude Code. The company has launched a public beta for self-hosted environments, allowing organizations to run Claude Code sessions on infrastructure they operate themselves rather than on Anthropic-hosted compute.

For companies experimenting with AI coding agents, that distinction is more meaningful than it may initially sound. The appeal of a tool like Claude Code is not merely that it can draft a function or explain an error message. Its real value comes when it can work near the systems developers actually use: private source repositories, internal package registries, CI pipelines, databases, proprietary command-line tools, and the sometimes messy web of services that make up a modern engineering organization.

That is also where security, governance, and network boundaries become non-negotiable. Anthropic’s new option is designed for teams that want Claude Code to operate in that environment without shifting the session’s working files and execution layer onto infrastructure managed by Anthropic.

The launch, announced August 6, arrives as AI coding assistants move beyond autocomplete and chat windows toward more agent-like work. Developers increasingly expect these tools to inspect a repository, run commands, test changes, troubleshoot failed builds, and produce pull requests. But granting an AI agent access to those capabilities can be a hard sell in organizations where code, secrets, and internal tools cannot freely cross a public network boundary.

Self-hosted environments are Anthropic’s answer to that problem. A developer can initiate a Claude Code session from the web, mobile app, desktop, or an automated routine, but the actual coding session runs inside the company’s network, alongside its internal services and under its existing security controls.

There is an important caveat, though. “Self-hosted” does not mean Claude becomes a fully on-premise AI system. Anthropic says repository checkouts, build artifacts, secrets, and files created or changed by a session remain on customer-provisioned infrastructure. But the conversation itself – including prompts, Claude’s responses, and tool results that may contain code Claude reads – is still sent to Anthropic for inference, while the transcript is stored so a session can continue across devices.

That split model will likely be the central point of scrutiny for security teams. It offers a practical middle ground: companies retain control of where the agent runs and where sensitive development artifacts live, without having to deploy or operate a frontier language model themselves. At the same time, organizations with rules that prohibit any relevant code or tool output from leaving their environment may find that this beta does not meet their requirements.

Anthropic is fairly direct about the target audience. The company says its standard hosted offering remains the recommended choice for most enterprises because it avoids the operational work of maintaining infrastructure. The self-hosted option is intended for organizations whose network design, internal tooling, or compliance obligations make customer-controlled execution necessary.

In practical terms, this is a product for platform engineering and developer-experience teams, not an easy switch that an individual developer flips on a Friday afternoon. Organizations need to provision and manage the underlying setup, including the runner image that powers sessions. They must also handle runner updates and, if they choose a dynamic deployment model, operate the orchestration layer that brings capacity online when sessions are queued.

Anthropic’s architecture revolves around long-running processes called runners. These runners pick up incoming Claude Code sessions and start a Claude Code process for each one. Companies can use a fixed model, where a set number of runners are always available and work is distributed among them, or an on-demand model, where an orchestrator detects queued work, starts runners as required, and stops them after the work finishes.

The distinction matters because AI coding agents can have unpredictable demand. A fixed pool may make sense for an engineering team with steady, daily use, while on-demand infrastructure may be more attractive for a large organization that wants to avoid keeping idle compute running. Anthropic says runners can serve multiple sessions, though every session gets its own repository checkout, helping isolate one developer’s work from another’s.

The setup also allows companies to prepare a standardized development environment before Claude begins work. Teams can preinstall compilers, software development kits, internal command-line interfaces, and other dependencies, so each new session starts with the same tooling developers would expect in a properly configured workstation or build environment.

That could be a bigger deal than it sounds. One of the recurring limitations of AI coding tools is the gap between a polished demo and a real production environment. In a demo, an agent might edit a simple app in a clean repository. In an enterprise, it may need a particular compiler version, access to a private artifact registry, credentials for a staging system, a custom build script, and knowledge of conventions that have accumulated over years. A self-hosted environment does not solve every part of that problem, but it gives teams a framework to place Claude Code much closer to the reality of their software stack.

Anthropic also positions the release as a way to connect Claude Code to internal workflows without exposing private services to the public internet. The company says preview users have adopted the setup for network access, customization, and compliance, including cases where Claude can help generate pull requests, investigate CI failures, and respond to events in developer workflows.

There is a clear difference between this feature and Claude Code’s Remote Control capability. Remote Control allows a developer to continue a Claude Code session that is running on their own machine from another device, such as a browser or phone. It remains tied to that individual’s machine and ends when that machine stops the session. Self-hosted environments, by contrast, run on shared infrastructure operated by the organization and can be used by any authorized user.

For enterprises, that shared-infrastructure approach is likely the more consequential one. It turns Claude Code from a personal developer tool into something closer to a centrally managed internal service. That can make adoption easier to govern, but it also raises the bar for implementation. Teams will need to decide which repositories and networks agents can access, how credentials are managed, what commands the agent may run, how activity is logged, and who is responsible when an automated change goes wrong.

The beta is available to organizations on Claude Team and Enterprise plans, but it is turned off by default. It is also unavailable to organizations using Anthropic’s Zero Data Retention, or ZDR, configuration. The limitation underscores the hybrid nature of the offering: while customers control the execution environment and its local artifacts, Anthropic still needs to process and retain session-related information to support the cross-surface experience described in the announcement.

Anthropic’s move reflects a broader shift in enterprise AI. The question is no longer just which model writes the best code. It is increasingly about where the tool runs, what it can reach, how it fits into established developer workflows, and whether a company can use it without weakening the controls it has built around its most valuable software assets.

For the right organization, self-hosted Claude Code environments could make AI agents feel less like an outside service being invited into the development process and more like another managed component of the engineering platform. The trade-off is straightforward: companies gain more control and deeper access to internal systems, but they also take on more responsibility for running the machinery behind it.