Self-hosted Claude Managed Agents: you run the server now

Anthropic will now let you run its agent service on your own servers. Self-hosted Claude Managed Agents launched on November 12. You host the API, the agent loop, the database and the sandboxes. Agents, sessions and session history stay with you. The model does not. Each call still goes out to a provider, and you take over the work Anthropic did before. Read the limits before you ask for access.

What actually happened

The release note is short: you run "the whole Managed Agents service, not only self-hosted sandboxes" on your own infrastructure. Access is by request through an Anthropic account team. No public pricing.

The docs give the parts list:

• Server. cma-server serves the API, runs the agent loop, and stores agents and history.
• Database. PostgreSQL 16 or later. With no database, the server keeps data in memory and a restart erases it.
• Workers. They run shell and file tools in your sandboxes.
• Production. Two Helm charts, Kubernetes 1.30 or later, gVisor nodes. Docker Compose is for development only.
• Model provider. The Claude API by default, or Amazon Bedrock, Google Cloud, Microsoft Foundry, or a gateway you run.

For each model call, the server sends the conversation so far, tool inputs and results included, to that provider. As The Next Web put it, Claude itself does not run inside your network.

The differences page is the one to read twice. Nothing migrates from the hosted service. There are no webhooks and no scheduled runs, so you bring your own cron. The prebuilt docx, pdf, pptx and xlsx skills are rejected. MCP servers, Git hosts and web tools are off until you allowlist them.

Why self-hosted Managed Agents matter for your business

This is a data-custody product, not a privacy shield. It answers one question well: who stores the session history and the credentials? You do. It also lets a company with a Bedrock or Foundry contract keep model spend on that contract.

It does not answer "does our data leave the building?" It does, on every model call.

And it costs you. You secure the server, patch it, install each release, and run Postgres and a Kubernetes cluster with sandbox isolation. For most small businesses, hosted agents with tight scopes are the correct answer.

Ask three questions before you request access:

1. What rule requires it? Name the contract clause or the regulation. "Feels safer" is not a rule.
2. Who patches it? Put a name next to the server, the database and the cluster.
3. What do you lose? Check your current agents against the differences page. Webhook triggers and document skills are the common breaks.

Sources: Claude Platform release notes, Self-hosted Claude Managed Agents docs, The Next Web.