Anthropic announced a self-hosted option for Claude Managed Agents on 12 November 2026, giving businesses control over the service that coordinates their AI agents and stores their work, as well as the tools those agents run.
Businesses can keep the agent service and stored conversations in their own cloud account or data centre, but model requests still go to a provider. This is an infrastructure option for teams building agents around internal documents and systems, not an offline version of Claude.
To request access, customers must contact their Anthropic account team. The company’s pricing page also directs self-hosted enquiries to that team, without listing a standard rate.
What businesses can bring in-house
Managed Agents coordinates Claude’s use of tools to carry out tasks. Anthropic’s earlier self-hosted sandboxes—isolated workspaces where those tools run—left the coordinating service and stored session history with Anthropic. The new option moves those parts to the customer too.
This is a different product from Claude Code’s self-hosted runners, covered previously by RohitAI; this announcement concerns the Managed Agents service.
Local history is not local model processing
Anthropic says each model request includes the conversation so far, including prompts and tool inputs and results. In practical terms, keeping a document on a company server does not prevent an excerpt from leaving it if a tool includes that excerpt in a request to an external model provider.
The default destination is Anthropic’s Claude API. Documented alternatives include Amazon Bedrock, Google-managed Claude Models on Google Cloud, Microsoft Foundry and a company-run gateway that routes model requests. The option does not supply a locally running Claude model.
The chosen provider’s data-retention terms still apply. Existing zero-data-retention arrangements cover eligible Claude API calls, not every model or feature. For example, Anthropic’s retention rules require 30 days for Fable 5.2 and Mythos 5.2 unless it authorises an exception.
The responsibilities and migration limits
Customers must run and secure the server, database and tool sandboxes, and install releases themselves. They also control how long their own histories are kept: deleting a session does not remove copies in database backups, exported logs or memory stores.
Moving an existing workflow involves more than changing its server address. Anthropic’s migration documentation says hosted resources and session history are not copied over; agents, skills, credentials and memory stores must be recreated. Scheduled deployments and webhooks—notifications used to trigger other software—are also absent, so workflows that depend on them need changes.
For a buying team, the useful starting point is to separate three needs: running tools on company infrastructure, storing conversations there, and choosing an approved model provider. Anthropic’s deployment guide says sandbox-only hosting is enough for the first; the full self-hosted service addresses the latter two.
A cost comparison should then include the quoted self-hosted price, model usage, servers, storage, backups and operator time. The published pricing information is not enough to establish a saving.