Skip to main content

Self-Hosted Claude Managed Agents: What Stays Inside Your Network

Self-hosted Claude Managed Agents, in private beta since 12 November, lets a company run Anthropic's agent server, its session database, and the sandboxes where tools execute on its own cloud or hardware. Session history and credentials stay inside the network, but each model call still sends the conversation, including tool results, to Anthropic or another chosen cloud provider.

On this page

What Anthropic launched

Anthropic opened a private beta of self-hosted Claude Managed Agents on 12 November, with access requested through a customer's Anthropic account team. Managed Agents is the company's service for running Claude as an autonomous agent that holds a session, calls tools, and works through a task over many steps.

Until now, customers could host only the sandbox, the container where an agent runs shell commands and edits files, while Anthropic ran the rest. The new option moves the whole service: the API, the agent loop, and the storage for agents, sessions, and session history. A customer runs a server, a PostgreSQL database, and workers that execute tools, and the API stays the same as the hosted version, so existing code and the command line tool keep working. Anthropic said teams that wanted this setup previously had to build the platform themselves.

What stays in your network

The privacy gain is real and specific. Session transcripts, which record every tool call and its result, now live in the customer's own database instead of on Anthropic's servers. That matters because Anthropic's own documentation lists the hosted Managed Agents service as not covered by zero data retention, precisely because Anthropic stores those transcripts until a customer deletes them.

Anthropic says it has no remote access to a self-hosted server or its log, and none of the credentials the server accepts are issued or held by Anthropic. Customers can keep tool secrets in their own secret manager, sign people in through their own identity provider, and restrict which MCP servers and Git hosts an agent may reach with explicit allowlists.

What still leaves

The model does not move. Claude still runs outside the customer's network, so prompts go out and completions come back. For every model call, the server sends the conversation so far, including tool inputs and tool results, to the inference provider. That provider can be Anthropic, Amazon Bedrock, Claude on Google Cloud, Microsoft Foundry, or a company's own gateway.

When Anthropic is the provider, each call also carries session, agent, and environment IDs in headers that cannot be turned off, plus a fixed label Anthropic uses to count self-hosted usage. The IDs hold no names or content, but they do encode when each was created. Customers with a zero data retention agreement get the same coverage for these calls as for other Messages API requests.

So the honest description is self-hosted orchestration, not self-hosted AI. Whatever an agent reads from your systems still reaches a cloud model in its context window.

The trade-offs for teams

Running the service yourself means securing a server, a database, and sandboxes, and installing each release. Anthropic warns that the agent's bash tool is a real shell with the worker's permissions, so isolation depends entirely on the container or virtual machine around it. Each tenant or trust level also needs its own deployment, because one server API key reaches every environment inside it.

Some hosted features are missing. There is no Files API, no webhooks, no scheduled deployments, and no prebuilt document skills, and web search and fetch are off until an operator enables them. With Anthropic as the provider, a request a safety classifier declines can also be answered silently by a different, possibly less capable Claude model unless that fallback is switched off.

Who this is for

For regulated companies that could not accept agent transcripts sitting with a vendor, this removes a real blocker. For anyone whose concern is the model seeing their data at all, it changes little: only a model that runs on your own hardware keeps every prompt local.

CuriousLM runs supported AI models locally on your device. Try CuriousLM.