Anthropic Opens Self-Hosted Claude Managed Agents in Private Beta

What happened

Anthropic's developer account announced the beta on X at 17:57 UTC on Nov 12, and Anthropic published an introductory article and documentation the same day. The beta is open only to eligible organizations, which apply through an Anthropic account team; Anthropic has not said which organizations qualify, and has published no pricing or general-availability date. The license calls the software grant royalty-free.

A self-hosted deployment has three parts: cma-server, which serves the same Managed Agents API as the hosted service, runs the agent loop and stores agents, sessions and session history; a PostgreSQL 16 or later database that holds that history; and workers that run an agent's shell and file tools inside the customer's sandboxes. The documentation says a Docker Compose setup is for development only and not for production, pointing teams toward Helm-based deployment; each deployment serves one tenant at one trust level.

The model itself does not move. Each model call goes to the Claude API by default, or to a provider the customer picks: Amazon Bedrock, Claude Models on Google Cloud, Microsoft Foundry or the customer's own model gateway. Anthropic's documentation says it has no remote access to a customer's self-hosted server or its logs, and session history stays in the customer's own database until the customer deletes it.

Differences from the hosted service

The self-hosted server lacks seven hosted features: cloud environments, the Messages API and Models API, the Files API, OAuth credentials for MCP servers, prebuilt document skills for docx, pdf, pptx and xlsx files, scheduled deployments and dreams, and webhooks. Web search and web fetch stay switched off until a team explicitly enables them. Nothing migrates from the hosted service, session history included, so customers rebuild agents, skills and history.

The Software License Agreement for the Claude Managed Agents Server, published the same day, requires a licensee to roll back a version or disable a feature within 8 hours of a written safety notice and to deploy a safety update within 1 business day; an agent incident must be reported to Anthropic within 48 hours of confirmation, and the agreement caps Anthropic's total liability at US$100,000.

Key facts

1. Launch: Self-hosted Claude Managed Agents entered private beta on Nov 12, open only to eligible organizations that apply through an Anthropic account team
2. Deployment: Three parts: cma-server, a PostgreSQL 16 or later database and workers; Docker Compose is for development only, with Helm-based deployment recommended for production
3. Model: Model calls go to the Claude API by default, or to Amazon Bedrock, Google Cloud, Microsoft Foundry or the customer's own gateway
4. Data: Agents, sessions and session history stay on the customer's own infrastructure; Anthropic says it has no remote access to a self-hosted server or its logs
5. Missing: Seven hosted features are missing, including cloud environments, the Files API, webhooks, scheduled deployments and dreams; web search and web fetch are off by default
6. License: Roll back within 8 hours of a safety notice and deploy a safety update within 1 business day; report incidents within 48 hours; liability capped at US$100,000
7. Pricing: No pricing or general-availability date was published; the license calls the software grant royalty-free
8. Background: Managed Agents launched as a fully hosted beta on April 8, 2026, gained self-hosted sandboxes in May; on Nov 6 Mistral also took Mistral Studio to customer VPC and on-premises deployments