Developer Tools

Anthropic Opens Self-Hosted Claude Managed Agents, Giving Enterprises Control of Agent Infrastructure

Anthropic has opened a private beta allowing organizations to run the full Claude Managed Agents infrastructure in their own cloud or on-premises environment. The move gives enterprise teams more control over agent credentials, session history and execution infrastructure, although requests to Claude still go to an external inference provider.

Conceptual editorial illustration: Private enterprise cloud infrastructure containing agent orchestration and credential vaults, with one clearly separated connection to an external AI model core

AI-generated editorial illustration by TeamRedevs; conceptual artwork.

Overview

Anthropic is expanding the deployment options for its autonomous AI-agent platform with a new self-hosted version of Claude Managed Agents. Announced on 12 November 2026, the private beta allows eligible organizations to operate the agent infrastructure within their own cloud accounts or on-premises data centers instead of relying entirely on Anthropic-hosted infrastructure.

The announcement addresses an important obstacle to enterprise AI adoption: companies increasingly want agents to work directly with internal code, databases and operational systems, but are cautious about where agent credentials, execution history and sensitive information are stored.

The new deployment model gives customers more control over these components without requiring them to build an entire agent platform from scratch.

What has changed?

Claude Managed Agents is Anthropic's collection of APIs and infrastructure for developing and operating long-running AI agents.

Previously, Anthropic operated the hosted service, while customers could already use self-hosted sandboxes for some workloads. The new beta moves more of the platform into the customer's environment.

Under the new architecture, organizations operate:

  • The Managed Agents server and control plane.
  • The database that stores agent session history.
  • Workers responsible for running shell and file tools.
  • The agent execution environment.
  • Credential storage and associated infrastructure.

Anthropic says developers can continue building agents with the same core APIs used by the hosted service.

This is important for businesses that want agents to perform tasks such as repairing failed builds, testing releases, investigating alerts and processing software backlogs without requiring employees to supervise every action.

Self-hosted infrastructure does not mean a self-hosted Claude model

There is an important distinction that enterprise buyers should understand.

The new deployment allows customers to host the agent service, session history, credentials and tool-execution infrastructure. It does not mean the Claude model itself is installed inside the company's network.

Model requests can be routed to:

Inference destinationPurpose
AnthropicDirect Claude model requests
Amazon BedrockModel access through AWS
Google's Agent PlatformModel access through Google's cloud platform
Microsoft FoundryModel access through Microsoft's platform
Customer-managed gatewayOrganization-controlled routing to its chosen provider

The agent's conversation and tool context still need to reach the selected inference provider when a model request is made. Anthropic says organizations with a zero-data-retention agreement receive the corresponding treatment for these requests under the same terms as other Messages API calls.

Therefore, companies should not interpret the announcement as a guarantee that all AI prompts or model-generated responses remain entirely inside their own infrastructure.

Four security principles behind the architecture

Anthropic describes four main design principles.

1. Separate the agent's reasoning loop from its execution environment

The agent harness, which coordinates model calls and tools, is kept separate from the sandbox where code runs. This prevents the execution environment from automatically inheriting access to every part of the agent's control infrastructure.

2. Keep credentials outside the sandbox

Anthropic describes a credential proxy that attaches credentials to supported MCP tool and API requests. Customers using the self-hosted deployment can also store secrets in their own secrets-management systems.

This is especially important when agents have permission to access private repositories, internal APIs or deployment services.

3. Restrict network destinations

Organizations can define approved hosts for each execution environment. Requests to destinations outside the allowlist are blocked. Customers operating their own workers must also configure the corresponding firewall protections correctly.

4. Maintain an audit trail

The platform records agent activity, including tool calls and their responses. These records can be exported through OpenTelemetry and integrated with existing monitoring and security systems.

For enterprise environments, this makes agent activity easier to investigate alongside conventional application and infrastructure logs.

What features are missing?

Self-hosting introduces additional operational responsibilities and some differences from the hosted service.

Anthropic's documentation identifies several limitations:

  • Cloud environments are not available in the self-hosted version.
  • The hosted Files API is not served by the self-hosted server.
  • Certain prebuilt document-processing skills are unavailable by default.
  • Scheduled deployments and Dreams are not supported in the self-hosted server.
  • The self-hosted server does not send webhooks.
  • Existing hosted agents, session history, credentials and memory stores are not automatically migrated.

Organizations must therefore recreate the relevant agents, skills, credentials and memory stores when moving to their own deployment.

This makes the new option more appropriate for teams that can operate infrastructure reliably, rather than businesses looking for a completely maintenance-free deployment.

What this means for developers and SaaS companies

For SaaS companies building AI-powered products, self-hosted Managed Agents could make it easier to meet requirements around internal data control, credential ownership and operational monitoring.

Engineering teams can consider using the platform for controlled workflows such as automated testing, issue investigation, build repair and internal software maintenance.

Before adopting it, teams should evaluate:

  1. Whether their compliance requirements demand self-hosted agent history.
  2. Which model provider will receive inference requests.
  3. How secrets and MCP credentials are protected.
  4. Whether network allowlists cover every required dependency.
  5. Who will maintain the server, database and worker infrastructure.
  6. How agent actions will be monitored and approved.

Availability and what to watch next

Self-hosted Claude Managed Agents is currently in private beta. Anthropic directs interested organizations to request access through their account teams. A general-availability date and separate self-hosted pricing were not established in the announcement and documentation reviewed for this briefing.

The next important question is whether enterprises will prefer fully hosted agent services, self-hosted orchestration or hybrid architectures that keep execution data inside their own infrastructure while using external model providers.

For developers, the announcement reinforces a larger industry trend: AI-agent infrastructure is becoming a distinct platform layer, and control over the harness, credentials, execution environment and audit history is increasingly part of the product decision.

Why this is trending

  • daily.dev: The self-hosted Managed Agents announcement appeared in the public Highlights feed approximately six hours before retrieval.
  • TechCrunch: The retrieved feed did not contain a matching dedicated report about this specific release.
  • X: Anthropic's developer announcement was referenced in reporting, but direct X access was unavailable for independent engagement verification.
  • Reddit: No reliable current topic-specific discussion was retrieved.
  • Hacker News: No matching current front-page discussion was identified.
  • Independent reporting: Help Net Security covered the release on 12 November, while The Next Web explained the distinction between hosting agent infrastructure and hosting the model itself.

The strongest signals are Anthropic's primary announcement, its official documentation, coverage by independent technology publications and its appearance in daily.dev's developer-news highlights.

Sources

  1. Anthropic — Operate Claude Managed Agents inside your security perimeter — 12 November 2026. Primary announcement explaining the deployment model, supported inference destinations and security architecture.
  2. Claude Platform Docs — Differences from the hosted service — Current documentation retrieved 13 November 2026. Technical limitations, feature differences and migration considerations.
  3. Help Net Security — Self-hosted Claude Managed Agents keeps agent credentials in your environment — 12 November 2026. Independent technical coverage of credentials, network controls and the deployment model.
  4. The Next Web — Anthropic lets companies run Claude's agent service on their own servers — 12 November 2026. Independent reporting clarifying that model calls still reach an external inference provider.
  5. daily.dev Highlights — Retrieved 13 November 2026. Developer-news feed showing the announcement among recent stories.
Share

Further reading

Building a product in this space?

MORE FROM THE NEWSROOM

Related news

All news
Back to all news

READY WHEN YOU ARE

Building in this space?

We help teams turn industry shifts into shipped products. Let’s talk about your roadmap.