Anthropic lets companies keep Claude agent histories on their own servers

Anthropic opened a private beta on November 12 that lets companies run the Claude Managed Agents service on their own infrastructure. Participating teams can put the service and its database in a cloud account they control or in their own data center. That gives them a place to retain agent settings and work histories alongside the internal systems the agents use. The server can send model requests to Anthropic or to another supported provider.

Companies could already host a Managed Agents sandbox, the isolated workspace where an agent runs commands and works with files. In that setup, the tools ran on the customer’s infrastructure while Anthropic operated the service that directed them and retained the session history. The new option moves that service and its storage into the customer’s environment as well. Anthropic opened the new deployment option as a private beta, with access through its account teams.

Where the work history lives

The company’s server accepts requests from applications and coordinates the exchanges between the model and the tools. A separate worker carries out file and shell operations in a sandbox and returns their results. The server uses those results as the agent continues working. This arrangement lets the organization house the database that records a session separately from the temporary workspace in which its tools run.

The database holds agent definitions, sessions and their histories. The documentation specifies PostgreSQL 16 or later for persistent storage. The server can also run with state held in memory, in which case a restart erases it. A company that needs to reopen a session days later or retain a record for review has to account for that storage when it sets up the service.

The self-hosted edition also lets teams manage the credentials agents use for connected tools. Anthropic says a proxy outside the sandbox attaches credentials to supported requests, and organizations can configure the server to use their own secrets manager. The edition has a web console at the company server’s address, where developers can open sessions after signing in through the organization’s identity provider.

What goes to the model provider

The data and access guide says the server sends the conversation so far to its inference provider each time it calls the model. That context includes tool inputs and results. If an agent reads an error from an internal system and asks the model what to investigate next, the returned error can become part of that request. The organization can keep the stored session on its own servers while another service processes that context.

Anthropic’s Claude API is the default destination. The documentation also lists Amazon Bedrock, managed Claude models on Google Cloud, Microsoft Foundry and a model gateway chosen by the customer. A gateway passes the request to the provider configured behind it. Anthropic says organizations with an existing zero-data-retention arrangement get the same treatment for requests sent directly to Claude as they do for other Messages API calls.

Developers can use the same core agent APIs, with some differences in the self-hosted edition. Existing agent definitions and session histories do not carry over automatically. Anthropic’s migration documentation says organizations must re-create agent settings, skills, vault credentials and memory stores on the new server. The self-hosted edition also leaves scheduled deployments and webhooks out. A company that runs a session at a set time can start it with its own scheduler.

The beta puts responsibility for running the server, storage and tool execution environments on the customer, including installing new releases. The company documents a single-machine development setup and a Kubernetes setup for production. Customers must contact their Anthropic account team for access to the installation package. Anthropic’s announcement gives no general-availability date or separate price for this deployment option.