Anthropic opens self-hosted Claude Managed Agents beta for clouds and on-premise data centers
Anthropic opened a private beta on 12 November 2026 that lets companies run the whole Claude Managed Agents service and its sandboxes on their own servers.
By Emre A.Published
Tech AI LatestAnthropic launched self-hosted Claude Managed Agents in private beta on 12 November 2026. Companies can now run the entire agent service, including the agent loop, sandboxes, credentials and session history, in their own cloud account or data center. Model calls still go to Anthropic by default, or to Amazon Bedrock, Google Cloud, Microsoft Foundry or a company's own gateway. Access is by request through Anthropic's account team; no price is published.
Key facts
- Announced: 12 November 2026 in the Claude Platform release notes and an Anthropic blog post
- Status: private beta; access is requested through the Anthropic account team
- Customer runs: the cma-server, a PostgreSQL 16 or later database, and the workers that run shell and file tools in sandboxes
- Model calls: Claude API by default, or Amazon Bedrock, Claude Models on Google Cloud, Microsoft Foundry or a customer gateway
- Missing versus hosted: webhooks, scheduled deployments, the Files API and prebuilt document skills, among others
- Not stated: pricing and a general availability date
What Anthropic launched
Anthropic has opened a private beta of self-hosted Claude Managed Agents, a version of its agent platform that companies run on their own infrastructure. The Claude Platform release notes date the launch to 12 November 2026, and an Anthropic blog post titled "Operate Claude Managed Agents inside your security perimeter" was published the same day. Access is not self-serve: companies ask their Anthropic account team for the release archive.
Claude Managed Agents is Anthropic's set of APIs for building autonomous agents. A developer defines an agent, starts a session and exchanges events with the agent while it works. Until now there were two ways to run it. In the default setup Anthropic hosts everything. With self-hosted sandboxes, introduced earlier, the containers where an agent's tools run commands and edit files move to the customer, while Anthropic still runs the service. The new option moves the service as well: the API, the agent loop and the storage for agents, sessions and session history.
Anthropic says the self-hosted server exposes the same API as the hosted service, so the Claude SDK and the ant command line tool work against it, apart from a documented list of differences.
What customers run and where Claude runs
According to the documentation, one deployment has four parts. A server called cma-server serves the API, runs the agent loop and stores data; it never runs an agent's shell or file tools itself. A PostgreSQL 16 or later database holds that data; without one, the server keeps everything in memory and a restart erases it. Workers pull jobs from the server and run tools inside sandboxes. A web console on the server lets staff sign in through the company's identity provider. For a first test, Anthropic ships a Docker Compose stack that it says is for development only because it serves plain HTTP, and recommends its two Helm charts for production on Kubernetes.
The model does not move. Each time the agent needs Claude, the server sends the conversation so far, including tool inputs and results, to an inference provider. By default that is the Claude API, but the server can instead call Amazon Bedrock, Claude Models on Google Cloud, Microsoft Foundry or a company's own model gateway. Session history stays in the customer's database, and Anthropic says that a zero data retention agreement applies to these model calls the same way it applies to any other Messages API request.
Security becomes the operator's job. The documentation warns that the agent's bash tool is a real shell with the worker's permissions and network access, and that isolation comes only from the container or virtual machine the worker runs in.
Why it matters and what is missing
Anthropic frames the launch around agents that now fix broken builds, test releases and triage alerts with access to internal systems. For banks, governments and other regulated buyers, keeping session logs, credentials and tool execution inside their own perimeter is often a condition for running such agents at all. The option also lets companies that buy Claude through a cloud marketplace keep agent traffic on that contract. Rivals are moving the same way: Mistral opened Mistral Studio managed agents in public preview, and Anthropic recently added inference hooks for Claude Enterprise.
The self-hosted server is not feature complete. Anthropic's list of differences says it has no webhooks, no scheduled deployments and no dreams, does not serve the Files API or the Messages API, rejects the prebuilt docx, pdf, pptx and xlsx skills, and limits multiagent orchestration to a coordinator with up to 20 listed subagents. Nothing is copied from the hosted service, so agents and session history must be recreated. Anthropic has not published a price or a date for general availability. More agent platform news is on the AI agents hub.
What changed
| Before | Managed Agents ran as Anthropic's hosted service; customers could host only the sandboxes |
|---|---|
| After | Customers can host the whole service, database and sandboxes, and route model calls to Bedrock, Google Cloud, Foundry or their own gateway |
| Who it affects | Enterprises with data residency or security rules that kept agent session history and credentials off third-party servers |
Questions people ask
- How do I get self-hosted Claude Managed Agents?
- It is in private beta. Anthropic says companies request access to the release archive through their Anthropic account team.
- Does the Claude model run on my servers?
- No. Only the agent service and sandboxes move. Model calls go to the Claude API by default, or to Amazon Bedrock, Claude Models on Google Cloud, Microsoft Foundry or your own gateway.
- What does self-hosted Claude Managed Agents cost?
- Anthropic has not published a price for the self-hosted server. Model usage is paid through whichever inference provider the server calls.
Sources
- Self-hosted Claude Managed Agents (Anthropic, 12 Nov 2026)
- Operate Claude Managed Agents inside your security perimeter (Anthropic, 12 Nov 2026)
- Claude Platform release notes (Anthropic, 12 Nov 2026)
- Differences from the hosted service (Anthropic, 12 Nov 2026)
- Get started with self-hosted Claude Managed Agents (Anthropic, 12 Nov 2026)
Reference pages
- Anthropic company
- Claude Managed Agents tool
- Amazon Bedrock tool
- Microsoft Foundry tool
Spotted an error? See the corrections policy and tell the editor. Facts in this story come from the sources listed above; vendor claims are attributed.
Comments
Loading comments.





