Claude Managed Agents Go Self-Hosted, With Pricing by Quote

Anthropic has opened a private beta that lets companies self-host the whole Claude Managed Agents service in their own cloud account or on-premise data center, it said in an announcement and documentation published on 12 November. Customers run the server that holds the agent loop, its database and the sandboxes where agents’ tools run. “The model runs at Anthropic or at a provider that you choose”, the docs say.

The hosted service lists a price: tokens at model rates plus “$0.08 USD per session-hour”, per Anthropic’s pricing page. For the self-hosted version, the same page says to “contact your Anthropic account team”, which is also how customers request access. The launch came 177 days after Anthropic’s self-hosted sandboxes on 19 May, by our count from its release notes.

Business Pill · The Model and the Harness

An evergreen Business Engineer explainer on the harness, the system around an AI model that makes same-model products differ. The video teaches the general idea only and does not discuss Anthropic’s launch.

What Moves to the Customer

The docs split Managed Agents into two parts a customer can host. Sandboxes are “The containers, VMs, or other isolated workspaces where an agent’s tools run commands and work on files”. The service is “The API, the agent loop, and the storage for agents, sessions, and session history”. Self-hosted sandboxes moved the first part to customers in May; the new option moves the second as well.

A deployment is a server, cma-server, a database on PostgreSQL 16 or later, and workers that run the agent’s shell and file tools in the customer’s sandboxes, the docs say. “The server never runs an agent’s shell or file tools”. Customers also “install each release yourself”, in the docs’ words, and session history is stored in the customer’s own database, according to the announcement.

The security work moves too. What Anthropic’s sandbox security model assigns to Anthropic “is your responsibility, including securing the server, hosting memory stores, and detecting unusual usage”, the docs say. Each trust level needs its own deployment.

Five Routes for Model Calls

By default the server sends model calls to the Claude API. The docs list three cloud alternatives, Amazon Bedrock, Claude Models on Google Cloud (Google managed) and Microsoft Foundry, plus a model gateway the customer runs: five routes in all, by our count. The announcement puts it as “Model requests go where you choose, whether that’s Anthropic directly, your cloud provider (Amazon Bedrock, Google’s Agent Platform, or Microsoft Foundry), or your own gateway”.

Data still leaves the network on each model call. “For each model call, the server sends the conversation so far, including tool inputs and tool results, to your inference provider”, the docs say. If a customer has zero data retention with Anthropic, it applies to those requests “in the same way as any other Messages API call”, per the announcement.

What the Self-Hosted Version Lacks

Anthropic’s differences page has three tables: four rows of features only the customer’s server has, seven rows of hosted features that “aren’t available”, and 14 rows of features that “behave differently”, by our count. The server-only features are a choice of inference provider, single sign-on, the customer’s own secret manager and agent identity.

The missing rows include cloud environments, the Files API, Anthropic’s prebuilt docx, pdf, pptx and xlsx skills, scheduled deployments, dreams and webhooks. Multi-agent work is narrower: “An agent can delegate only to subagents that you list. Dynamic workflows, the advisor, and subagents that the agent defines itself aren’t available”. Those dynamic workflows are the feature we covered in our piece on workflows of up to 1,000 agents.

Google’s Sandbox Option, a Day Earlier

On Wednesday, 11 November, Google’s Gemini API managed agents gained self-hosted sandboxes, as we reported in our piece on Gemini agents’ sandboxes. In our reading of Google’s release, the agent loop and the model stay on Google’s side. Anthropic shipped its own self-hosted sandboxes on 19 May, 176 days before Google’s option, by our count, and now adds the service.

The beta lands as Anthropic sells shares in its initial public offering; we covered the European retail tranche in our piece on Europe’s IPO slice. Neither the announcement nor the docs we read mention the offering.

The Structural Read

As we read it, the move is aimed at buyers whose rules stop session history from leaving their network; the docs still send each model call, with tool inputs and results, to the chosen provider.

As we read the differences page, the trade has a cost in features: seven rows of hosted features are not available, including webhooks, scheduled deployments and dynamic workflows, by our count.

Three Implications

FOR ENTERPRISE BUYERS Access and self-hosted pricing both go through the Anthropic account team, the docs and pricing page say; the hosted service lists $0.08 per session-hour plus tokens.

FOR CLOUD PROVIDERS Amazon Bedrock, Claude Models on Google Cloud and Microsoft Foundry are named routes for the self-hosted server’s model calls, the docs say.

FOR RIVAL AGENT PLATFORMS Google’s Gemini API added self-hosted sandboxes on 11 November, as we reported; Anthropic’s new option also moves the agent loop and storage, the docs say.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Agentic Harness War.

The framework argues: “The agentic surface — the harness — is becoming where value pools in the entire AI stack.”

As we read it, letting customers run the harness themselves, while model calls can go to other clouds, keeps Anthropic’s harness in accounts that host their own stack; the framework predates the launch and does not discuss it.

What Is Not Established

The pages we read do not give a price for the self-hosted version, a date for general availability, or the names of beta customers. Anthropic’s announcement describes the security design and the docs describe the setup; neither gives usage figures for Managed Agents. Anthropic was not contacted for this piece.

The Bottom Line

Anthropic now lets companies run the whole Claude Managed Agents service, not only its sandboxes, on their own infrastructure, in a private beta priced through its account team. Model calls can go to the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry or a customer’s gateway, and seven rows of hosted features are not available on the self-hosted server, by our count.