Anthropic Opens Self-Hosted Control Plane for Claude Agents in Private Beta
Session history and credentials move to customer infrastructure, but model prompts still leave the network.

Anthropic has opened a private beta for self-hosted Claude Managed Agents, moving the agent orchestration layer — the server that runs the agent loop, stores session history, and hands work to tool-running workers — onto enterprise customers' own infrastructure for the first time. The company announced the option on November 12, extending a data-residency effort that began in May, when customers were first allowed to run only the execution sandbox on their own hardware. What the change does not eliminate is the most consequential data crossing: model inference. The prompts agents send to Claude, and the completions they receive, still travel to Anthropic or to another model provider the customer selects.
What Moved Inside the Perimeter and What Did Not
Before this week, teams running Claude Managed Agents in tightly controlled environments faced an architectural limit. Even after the May 2026 rollout of self-hosted sandboxes, the isolated environments where an agent's code runs, Anthropic operated everything else, including credential handling and the record of what each agent did, according to Help Net Security. That is where the operationally sensitive data accumulates: the session history recording every agent action and the logic deciding which tool to call next. In its announcement, Anthropic said some teams want the whole service in their own environment, adding: "Until now, that meant building the platform themselves."
The private beta shifts three components to the customer, which can run them in a cloud account or an on-premise data center. Anthropic's documentation for the self-hosted service describes a server program named cma-server that serves the API, runs the agent loop, and stores agents, sessions, and each session's history. A PostgreSQL database, version 16 or later, provides persistent storage; without one, the server keeps its data in memory and a restart erases it. Workers, started with the ant beta:worker command or built with Anthropic's SDK, take work from the server and run the agent's shell and file tools in sandboxes the customer defines. The workers poll the server for jobs, and the documentation says the server never runs those tools itself.
The security model Anthropic built for the hosted service carries over. The harness that calls Claude is kept separate from the sandbox where agent code runs, so the agent loop, session history, and stored secrets stay outside the sandbox. For MCP tools and what Anthropic calls supported API calls, a proxy outside the sandbox attaches the credential to each request, and self-hosted customers can keep those credentials in their own secrets manager. Help Net Security noted that the protection is limited to those two categories, so any other service an agent reaches deserves separate review. Each environment can be limited to an allowlist of hosts, and the record of every tool call can be exported over OpenTelemetry to existing monitoring tools. Anthropic's data and access documentation states: "Anthropic has no remote access to the server or its log." That is a company statement, not an independently audited finding.
One isolation duty moves with the software. The documentation warns that the agent's bash tool is a real shell with the worker process's permissions and network access, and that the worker does not isolate it. The strength of the sandbox therefore depends on the container or virtual machine the customer builds around each worker.
Five Paths for Model Inference, None Fully On-Premise
The portion that always leaves the server is model inference. For each model call, the server sends the conversation so far, including tool inputs and tool results, to an inference provider. Anthropic's inference provider documentation lists five settings: the Claude API, which is the default; Amazon Bedrock; Claude models on Google Cloud; Microsoft Foundry; and a model gateway the customer runs. The gateway option lets an organization route every model request through its own infrastructure, where it can centralize credentials and access rules, but the gateway forwards those requests to a model that runs elsewhere. Anthropic says the server connects to Anthropic only when the Claude API is the chosen provider.
The routing choice also affects features. According to Anthropic's page on differences from the hosted service, the web search tool works when the provider is the Claude API, Google Cloud, or Microsoft Foundry, and the web fetch tool works only with the Claude API and Foundry. Amazon Bedrock appears on neither list, and both tools stay off until an administrator enables them. Attribution differs too. When the Claude API is the provider, each model call carries up to three headers holding the session, agent, and environment IDs, and no setting turns them off. A separate label identifying the server software is sent to the Claude API, Microsoft Foundry, and compatible gateways, but never to Amazon Bedrock or Google Cloud.
Zero Data Retention Applies, but HIPAA Coverage Does Not
Self-hosting changes one compliance answer and leaves another untouched. Anthropic's Managed Agents overview explains that the platform is stateful by design: sessions are long-running, and transcripts persist until the customer deletes them. For that reason the hosted service is not eligible for Zero Data Retention, the arrangement under which Anthropic does not store API request data. The documentation now makes an exception for the self-hosted service, where the customer stores that data. If an organization has a zero data retention arrangement with Anthropic, it covers the self-hosted server's model calls in the same way as other Messages API calls.
No comparable exception appears for healthcare data. The same overview lists Managed Agents as not eligible for coverage under a HIPAA Business Associate Agreement and says that status applies to all Managed Agents sub-features. Anthropic's announcement of the self-hosted option does not mention HIPAA. The practical reading is that organizations handling Protected Health Information should not assume that moving the server and database in-house brings agent sessions under Anthropic's BAA. Anthropic's documentation adds that model inputs and outputs fall under the organization's agreement with whichever model provider it selects, which makes the provider contract a second place where healthcare teams need a clear answer.
For other regulated sectors the picture is more favorable. Session history sits in the customer's own PostgreSQL database until the customer deletes it, and the server does not delete sessions based on age. Deleting a session through the API does not reach database backups or exported logs and traces, so retention policy becomes the customer's job in full.
How the Large Cloud Agent Platforms Compare
The agent platforms sold by the three largest cloud providers take a different approach to the same problem. Amazon Web Services describes Bedrock AgentCore as HIPAA eligible, and Google Cloud says its Vertex AI Agent Engine supports HIPAA workloads on a managed runtime. Microsoft's Foundry Agent Service offers a standard setup that stores agent state in the customer's own Azure resources, which narrows the gap on where conversation data lives.
Each of those services is documented as running inside its vendor's cloud. Anthropic's self-hosted option differs in that the orchestration server itself is software the customer installs and operates, including in an on-premise data center, without building an agent platform from scratch. The trade-off runs both ways: the cloud providers pair a vendor-run control plane with HIPAA eligibility, while Anthropic pairs a customer-run control plane with no HIPAA coverage for the product.
Operational Obligations Enterprises Should Understand Before Deploying
Anthropic's documentation says the server software is licensed under a dedicated Software License Agreement for the Claude Managed Agents Server. According to a review of that license text published by CellCog, the agreement took effect on November 12 and carries obligations that go beyond standard API terms. Customers must deploy a safety update within one business day. After a written safety notice, they have eight hours to disable a feature or roll back a version. Other updates carry a 30-day window, an agent incident must be reported within 48 hours of being confirmed, and unauthorized access to the code must be reported within 72 hours. The review puts the agreement's total liability cap at US$100,000 and the notice period for termination without cause at 90 days. CellCog noted it could not confirm that the published text is the version beta customers sign.
The same review says the license permits the software to send usage signals to Anthropic and bars licensees from removing, blocking, altering, or falsifying them. Anthropic's documentation describes what those signals are in practice: the ID headers and software label attached to model calls, which contain server-generated identifiers and no names or content. Organizations in highly restricted environments will need to decide whether that outbound metadata is acceptable under their own security policies. The license also allows Anthropic to limit certain models to its hosted service for safety reasons, according to the review.
What Remains Hosted-Only
Several features of the hosted service are absent from the self-hosted server. Anthropic-managed cloud environments do not exist there; every environment must be backed by the customer's own workers. The Files API is unavailable, so files have to be placed in the worker's working directory. The server sends no webhooks and has no route for scheduled deployments, which leaves scheduling to the customer's own tooling. OAuth-based credentials for MCP servers are rejected because the server has no flow to refresh tokens, and the prebuilt skills for Word, PDF, PowerPoint, and Excel files must be uploaded as custom skills.
Memory is a partial exception. Memory stores work on the self-hosted server, with the worker rather than the server copying each store into the sandbox and writing changes back. Dreams, a research-preview feature in which Claude reviews past sessions to curate an agent's memory, do not run in self-hosted mode, and Anthropic's documentation says they have no replacement. Teams that rely on automated memory curation will find that gap still open.
From Sandbox to Full Service: Seven Months of Architecture Decisions
Anthropic launched the hosted Managed Agents platform in public beta on April 8, 2026, naming early adopters including Notion, Rakuten, Asana, Sentry, and Vibecode. Self-hosted sandboxes followed in public beta on May 19, announced at the company's Code with Claude event in London alongside a research preview of MCP tunnels, which connect agents to private MCP servers without exposing them to the public internet. Scheduled deployments entered public beta on June 9. The November 12 private beta for the full self-hosted service arrives seven months after the initial launch and moves the last Anthropic-operated layer other than the model itself behind the customer perimeter.
Anthropic has not announced general availability, and its announcement disclosed no pricing for the server software; access is by request through an Anthropic account team. Customers admitted to the private beta will be the first to test whether the reported update and incident-reporting deadlines fit enterprise change-management processes, which seldom move on a one-business-day schedule. That operational fit, as much as the architecture, is likely to shape how far self-hosted Claude Managed Agents reaches into the regulated industries it targets.



