Self-Hosted Claude Code: Power Inside the Firewall
Self-hosted Claude Code is Anthropic’s deployment option that runs coding agent sessions on infrastructure controlled by an organization, keeping repository checkouts, build artifacts, secrets, and files on in-house machines while sending conversational data and tool results to Anthropic’s API for model inference and transcript storage so work can continue across devices. Anthropic has moved these self-hosted environments for Claude Code into public beta, available to Team and Enterprise customers and announced on August 6, 2026. This is not a cosmetic change—it shifts where AI development happens and who owns the operational risk. With self-hosted Claude Code, sessions started from the web, mobile, desktop, the terminal, or scheduled routines now run inside the organization’s network, next to internal services, toolchains, and security controls. The coding agent becomes an on-premise coding agent rather than a distant cloud helper, and enterprises gain real data residency and network isolation—but in exchange, they inherit the duties traditionally shouldered by an AI provider.
What Moves On-Premise—and What Still Belongs to Anthropic
The core shift in enterprise AI infrastructure here is architectural, not rhetorical. Repository checkouts, build artifacts, secrets, and any files a Claude Code session creates or modifies now stay on machines provisioned by the organization. This is the data residency win: source code and outputs remain in-house, satisfying stricter compliance teams that refuse to let build pipelines sprawl across the public internet. But self-hosted does not mean fully self-contained. The conversation—prompts, responses, and tool results that may contain code—still flows to Anthropic’s systems for model inference, and session transcripts are stored so developers can resume work from web, mobile, or desktop apps. Every connection between the environment and Anthropic is outbound HTTPS to api.anthropic.com, covering queue polling, event streams, and inference, and Anthropic never connects into the customer’s network. That design keeps the model centralized while execution and integrations live behind your firewall, a deliberate hybrid rather than classic on-prem LLM deployment.
How Self-Hosted Claude Code Changes Day-to-Day Development
For ordinary developers, the change is subtle in the interface but significant in capability. A session started from the web, mobile, desktop, the terminal, or a scheduled routine now runs inside the organization’s network, adjacent to internal databases, registries, and services, without exposing them to the public internet. Teams can pre-install compilers, SDKs, and internal command-line tools so every session begins ready to build against their stack. This is where on-premise coding agents stop being novelty and start being workflow-critical. Anthropic says preview customers adopted the option because sessions can reach internal services without opening new public endpoints, development environments start pre-loaded with the right tooling, and source code plus build artifacts stay in-house for compliance. As George Jacob, a senior engineering manager, put it, “Self-hosted environments let us integrate Claude Code into our existing development workflows while maintaining our security and operational controls.” In practice, Claude can now generate PRs, help fix CI issues, and respond to developer workflow events with compute that scales based on demand.
Operational Burden: Runners, Queues, and Platform Ownership
The catch is straightforward: Anthropic is moving operations into your lap. An organization must create named environments in its admin settings, each grouping a set of runners—long-lived processes deployed on hosts inside its network that behave much like self-hosted CI runners. When a developer starts a cloud session and picks an environment, Anthropic’s control plane places the session on that environment’s queue, a runner with free capacity claims it, clones the chosen repository using provided credentials, and starts a Claude Code process on the host. Runners come in fixed mode, which keeps a set number running and distributes sessions, or on-demand mode, where an orchestrator starts runners as sessions arrive and stops them when work finishes so capacity tracks demand. A runner serves one user at a time, locking to that account to prevent code from mixing between users, and if a runner stops polling for roughly 60 seconds, the server requeues the session to another runner. Anthropic advises customers to assign a platform, developer experience, or developer productivity team to build and maintain the runner image, update runners, and operate the fleet; what changes hands is the operations.
Choosing Between Hosted Convenience and On-Premise Control
Anthropic is candid about its Claude deployment options: it recommends the hosted service for most enterprises because there is no infrastructure to operate or maintain. Self-hosted Claude Code is positioned for organizations whose network access, tooling, or compliance demands require greater control over where agent sessions execute. In other words, this is an opt-in trade: cloud convenience versus on-premise control. Hosted Claude removes infrastructure worries but keeps your coding agent at arm’s length from internal systems. Self-hosting lets sessions reach private services and keeps code and artifacts in-house, but brings an operational burden that not every company is ready to shoulder. Remote Control—a separate feature that allows a developer to continue a session running on their own machine from a phone or browser—remains distinct, tied to the user and ending when that machine stops, whereas self-hosted environments run on shared infrastructure managed by a platform team and serve many users. Notably, sessions from Claude Tag, Claude Security, and Code Review do not route to self-hosted environments yet; support for those surfaces will follow separately. Enterprises should see this beta not as a checkbox feature, but as a strategic fork in their AI infrastructure roadmap.
