Anthropic opens self-hosted Claude Managed Agents in private beta

Available on 12 November 2026 through Anthropic account teams, the private beta lets enterprises run the Managed Agents server, PostgreSQL state store and tool sandboxes inside their own cloud or on-prem network.

Anthropic launched self-hosted Claude Managed Agents in private beta on 12 November 2026, documented on Claude Platform Docs and announced by @ClaudeDevs. Whereas Anthropic’s earlier self-hosted sandboxes option kept the orchestration API on Anthropic infrastructure, and one day after Google added self-hosted sandboxes for Gemini Managed Agents, Anthropic’s new deployment mode moves the entire Managed Agents service into a customer’s own cloud account or on-premises data centre.

Control-plane architecture, PostgreSQL storage and provider routing

A self-hosted deployment comprises the cma-server binary (running the REST API and autonomous agent loop without executing shell or file tools itself), a PostgreSQL 16 or newer database persisting agents, sessions and full event history, and isolated workers launched via Docker Compose, the cma-sandboxes Helm chart, the ant beta:worker CLI or SDK workers. Because the server exposes the identical API as Anthropic’s hosted service, existing SDK code and the ant CLI connect without code rewrites, alongside a local web console wired to the customer’s identity provider and a cma-server-admin CLI for Linux (x86-64 and Arm) and Apple silicon macOS.

Administrators split trust across three separate keys (Server API key, Inference credential and per-queue Environment key) and route model calls to the Claude API by default or to Amazon Bedrock, Google-managed Claude Models on Google Cloud, Microsoft Foundry, or an internal model gateway. Enterprises must request binary download access through their Anthropic account team under the Claude Managed Agents Server licence; Anthropic has not published self-hosted licensing fees.

Technical details

Deployment components

cma-server control plane, PostgreSQL 16+ state store, local IdP web console and isolated sandbox workers

Supported inference endpoints

Claude API (default), Amazon Bedrock, Google Cloud (Google managed), Microsoft Foundry or custom model gateway

Packaging and admin tooling

Docker Compose stack, cma-sandboxes Helm chart, ant CLI and cma-server-admin (Linux x86-64/Arm, macOS Arm)

Access model

Private beta archive with operator-guide.md via Anthropic account teams (commercial server licence)

Release details & context

Security boundaries, multitenancy and outbound network traffic

Anthropic’s operator documentation stresses that an agent’s bash tool inherits the permissions and network reach of its worker process rather than isolating commands inside the worker itself, so production setups rely on the cma-sandboxes Kubernetes Helm chart or disposable VMs to start each session in a dedicated pod or container, with separate server deployments recommended per trust boundary.

Only model inference requests leave the customer’s network, carrying conversation turns, tool inputs and outputs, and when pointed at the default Claude API, session, agent and environment identifiers. Each release bundle ships with a versioned operator-guide.md file designed for Claude itself to read when assisting engineers with server setup.

Common questions

How does self-hosted Claude Managed Agents differ from self-hosted sandboxes?

Self-hosted sandboxes only run tool containers on your hardware while Anthropic hosts the control plane and history, whereas self-hosted Managed Agents moves cma-server, the API and PostgreSQL session storage onto your infrastructure too.

Which model providers can a self-hosted cma-server send Claude inference requests to?

Operators can point the server’s inference credential at the Claude API, Amazon Bedrock, Google-managed Claude Models on Google Cloud, Microsoft Foundry or an internal enterprise model gateway.

Anthropic release timeline (37 launches tracked)

Earlier6 Nov 2026

Anthropic Science Blog details how AI agents build and constrain scientific models

Anthropic publishes a Science Blog post by University of Tübingen researchers on combining simulation-based inference with AI agents for scientific models.

Earlier9 Nov 2026

Anthropic releases Claude Fable 5.2 generally and gated Claude Mythos 5.2 at lower API prices

Anthropic ships Claude Fable 5.2 and gated Mythos 5.2 on one shared model at $8 and $40 per million tokens, taking first place on Artificial Analysis v5 at 65.

Earlier10 Nov 2026

Anthropic launches Claude Whiteboards, Sheets and generally available Claude Data

Anthropic rolls out multiplayer Claude Whiteboards and Sheets in beta alongside generally available Claude Data across paid plans.