Independent · Updated November 20, 2026 · Every claim traceable
Home › AI Infrastructure › Self-Hosted Claude Managed Agents: What It Changes…
AI Infrastructure

Self-Hosted Claude Managed Agents: What It Changes

By Yogesh · · 8 min read

Sources checked and dated on publication. See our editorial policy.

Anthropic announced self-hosted Claude Managed Agents on November 12, 2026, opening private beta for teams that want…

Anthropic announced self-hosted Claude Managed Agents on November 12, 2026, opening private beta for teams that want the entire agent service inside their own perimeter. The agent harness, sandboxes, credentials and session history can all run in your own cloud or on-premise, while model requests go wherever you choose. What moves to your servers is exactly the part that used to be Anthropic's, and that is also where the operational work lands.

The short answer

What self-hosted Managed Agents covers

Anthropic first launched Managed Agents as a hosted service on the Claude Platform, where Anthropic runs the infrastructure: the agent harness, sandboxes, credential handling, and session history needed to get long-running agents into production. It later added self-hosted sandboxes, which let customers run tool execution on infrastructure they control while the agent loop stays with Anthropic.

Self-hosted Managed Agents goes further. According to Anthropic, every part of the service runs in your own cloud or on-premise, including the harness, sandboxes, credentials, and session history. You operate three concrete components: the Managed Agents server, its PostgreSQL database, and the workers where agents run their shell and file tools. Session history is stored in that database.

The developer experience is deliberately unchanged. Anthropic says developers still build agents with the same core APIs as the hosted Managed Agents service, so moving from hosted to self-hosted is a deployment decision rather than a rewrite. If you already have agents built against the hosted API, that compatibility is the main practical argument for the beta.

Model calls: you choose the endpoint

The other half of the architecture is where the model request goes. Anthropic says model requests can be routed to Anthropic directly, to your cloud provider - Amazon Bedrock, Google's Agent Platform, or Microsoft Foundry - or to your own gateway. This is the piece most often misunderstood, so it is worth stating plainly: self-hosting the agent service does not mean running the model locally, and it does not mean the request never leaves your network.

What you gain is control over the endpoint. If your data-protection posture requires traffic to leave through a specific cloud provider's channel or an internal gateway that enforces policy and logs, you can route it there. What you do not gain automatically is a data-retention exemption: Anthropic states that if your organization has zero data retention in place with Anthropic, it applies to requests sent to Claude in the same way as any other Messages API call. Self-hosting the harness does not create a ZDR agreement for requests that go to Anthropic, and it does not by itself make the model inference local.

The security design, and why it matters operationally

Anthropic frames agent security as an architectural property rather than a policy document, and it lists four principles that apply whether Anthropic hosts everything, you host only the sandbox, or you host the whole service.

The brain and the hands are split. The harness (the brain) calls Claude, while the sandbox (the hands) is where the agent's code runs. The agent's code executes in the sandbox, while the agent loop, session history, and vault secrets stay outside it. Anthropic says this separation means an agent does not depend on one sandbox surviving an entire run, that sandboxes start only when a tool needs them, and that session history lives outside the sandbox as an event log. In the self-hosted model, the same split is what lets you keep credentials and transcripts out of the environment where untrusted code executes.

Credentials stay out of the sandbox. Vaults are designed so that a proxy outside the sandbox adds the credential to requests for MCP tools and supported API calls. With self-hosted Managed Agents, Anthropic adds the option of storing credentials in your own secrets manager - which is the piece most teams will care about, since agent credentials usually have broad access to internal systems.

Network access can be limited. Each environment can be restricted to a list of allowed hosts, with requests to any other host blocked. Anthropic's example for an agent that fixes broken builds is a list containing just the git host and the package registry. When you host the sandbox yourself, you set the same limits in your own firewall - which usually gives you tighter control, because the enforcement point is inside your network.

Every session leaves a full record. Managed Agents keeps a record of every step an agent took, including the tool calls it made and what came back. That record exports over OpenTelemetry, so it can be sent to monitoring tools you already use and reviewed alongside the rest of your systems. For a self-hosted deployment this is a meaningful gain: the audit trail is in your own telemetry pipeline rather than a vendor console, and you control retention and access to it.

What you take on

Self-hosting moves a class of work that was previously invisible to most customers: running a server, a database and workers that must be available, patched, backed up, monitored and scaled. Anthropic is explicit that the customer operates all three components. If your team has never run a stateful service with production availability requirements, that is the real cost of this option, and it deserves to be weighed against the risk it removes.

Consider the failure modes. A managed agent working unattended on a codebase needs its session history to survive restarts, because that log is what a human reviews afterward to understand what the agent did. If the database is undersized, misconfigured, or restored from a stale backup, you lose the audit trail at exactly the moment you need it. Workers that run shell and file tools need isolation, patching and capacity planning like any other workload in your environment. If the git host or package registry allowlist is too tight, agents fail mid-task; too loose, and you have given an autonomous process wider network reach than you intended.

There is also a versioning question that the announcement does not answer directly: how do you run the same agent service version across development, staging and production, and how do you roll forward when Anthropic updates the service? Treat that as an open item to resolve with your account team during beta, along with the operational support boundary, since a private beta typically ships with a narrower support envelope than a generally available product.

Who should care, and who should wait

The fit is narrow but real. If your agents touch production code, internal ticketing, customer data pipelines or regulated records, and your security team requires the agent loop, credentials and transcript inside your perimeter, this release addresses a genuine blocker that previously meant building the platform yourself. If your agents only read public data and draft text, hosting the whole service is operational cost without a matching control benefit, and the hosted service is the better choice.

Note also that the announcement is scoped to eligible organizations requesting access through an Anthropic account team. A private beta means limited availability, evolving behaviour and no published pricing in the announcement, so this is a conversation to start now and a dependency not to plan a quarter around.

If you are still deciding whether an agent needs a sandbox at all, our guide to Google ADK 2.14 and A2A agent teams covers how frameworks split orchestration from execution, and our comparison of VPS options for agent workloads is useful if you end up running the workers yourself.

A practical evaluation checklist

If your organization is a good fit, evaluate the beta with a specific list rather than a general demo. First, map what runs where: the harness, the sandboxes, the credential store, the session database, and every model endpoint. Confirm with Anthropic in writing that your intended configuration is supported. Second, test the failure paths that matter to you - restore a session database from backup, restart a worker mid-run, revoke a credential, and block an allowlisted host - and confirm the behaviour is what your runbook assumes.

Third, verify the audit record end to end: check that OpenTelemetry export captures the tool calls and results, that the export lands in your own monitoring stack, and that retention matches your policy. Fourth, decide the model endpoint deliberately, including what happens to prompts if the endpoint is Anthropic rather than your cloud provider, and whether existing data-protection agreements cover it.

Fifth, size the operational cost. Measure the resources the server, database and workers actually consume under a realistic workload with parallel sessions, and compare that to the headcount cost of running the same thing today with a hosted agent service. That comparison, not the compliance checkbox, is what determines whether self-hosting is the right operating model for your team.

The bottom line

Self-hosted Claude Managed Agents closes a real gap. Teams that need agent credentials, transcripts and execution inside their own perimeter no longer have to choose between that requirement and using Anthropic's agent stack. The architecture is sensible - harness separated from sandbox, credentials outside the sandbox, allowlisted network, exported audit trail - and the API compatibility means existing agent builds can move over.

What you accept is responsibility for the infrastructure underneath: a server, a PostgreSQL database, workers running untrusted-adjacent code, and an audit trail whose integrity depends on your own backups and monitoring. For a narrow class of regulated or security-constrained teams, that is a good trade. For everyone else, the hosted service remains the sensible default until the beta matures. Request access, run one real workload through it, and judge the design on your own failure modes rather than on the announcement.

Frequently asked questions

Private beta, request access

Anthropic says self-hosted Claude Managed Agents is available in private beta and access is requested through an Anthropic account team.

You operate three components

The customer runs the Managed Agents server, its PostgreSQL database, and the workers where agents run shell and file tools; session history lives in that database.

Model routing stays yours to choose

Model requests can go to Anthropic directly, to Amazon Bedrock, Google’s Agent Platform, Microsoft Foundry, or to your own gateway.

Self-hosting is not zero data retention

Anthropic says zero data retention applies to requests sent to Claude only if the organization already has a ZDR agreement with Anthropic; hosting the service does not create one.

Sources

Prices and benchmark figures are standard published rates as of November 14, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.