Self-Hosted Claude Managed Agents Private Beta Explained: Anthropic Lets Companies Run AI Agents on Their Own Servers

Image: Anthropic — Claude logo (official brand mark). Source: Wikimedia Commons. The announcement came as a developer post with no product screenshot, so the logo stands in for the official visual.

TL;DR: The self-hosted Claude Managed Agents private beta, announced 12 November 2026, lets companies run Anthropic's full agent service on their own cloud or servers — the agent harness, sandboxes, credentials and session history all move inside their own perimeter. The model itself does not: every call to Claude still goes to Anthropic, Amazon Bedrock, Google, Microsoft Foundry or the company's own gateway.

📌 Key Highlights

• Anthropic opened a private beta on 12 November 2026 for fully self-hosted Claude Managed Agents; access runs through an Anthropic account team.
• Customers run three pieces themselves: the cma-server, a PostgreSQL 16+ database for session history, and the workers where agents run shell and file tools.
• The same Managed Agents API as the hosted service — the Claude SDK and the ant CLI connect to the customer's own server.
• The model stays outside: each request to Claude goes to a provider the customer chooses, so prompts still leave the network.
• Not everything moves over: scheduled deployments, dreams and webhooks are missing, and nothing migrates from the hosted service. No pricing or general-availability date has been published.

Self-Hosted Claude Managed Agents: What Moves Onto Your Servers

Claude Managed Agents is Anthropic's hosted way to run Claude as an autonomous agent: you define an agent once — model, instructions, tools and skills — and it keeps working in its own sandbox, fixing builds, testing releases and triaging alerts without anyone watching. Until now, that whole service ran on Anthropic's servers. In May 2026 Anthropic added self-hosted sandboxes, which let the code an agent runs execute on customer hardware — but the agent itself, its credentials and its history still lived with Anthropic.

The 12 November announcement moves the rest of it home. The customer operates the Managed Agents server (cma-server), which serves the API, runs the agent loop and stores agents, sessions and their histories; a PostgreSQL database, version 16 or later, that holds that history — without it, the server keeps data only in memory and a restart wipes it; and the workers that run the agent's shell and file tools inside the customer's sandboxes. The server itself never executes an agent's shell or file tools, which keeps the control plane separated from whatever the agent is doing. The self-hosted server exposes the same Managed Agents API as the hosted service, so the Claude SDK and the ant command-line tool work against a company's own server without new code. Anthropic says it has no remote access to a customer's self-hosted server or its logs.

The Model Still Lives Outside — and That Is the Point to Understand

The headline can mislead, so it is worth stating plainly: self-hosting the service does not keep model traffic inside the customer's network. Claude itself still runs outside. Every time an agent asks the model something, the conversation so far — prompts, tool inputs and tool results — travels to the inference provider the customer picks. Anthropic is the default, but the supported options are Amazon Bedrock, Google's Agent Platform and Microsoft Foundry, or the customer's own model gateway, set with the inference.provider setting.

What self-hosting does keep inside is everything around the model: the credentials agents use to reach internal systems, the record of every session, and the sandboxes where code runs. Anthropic built Managed Agents around four security principles that carry over here: the harness that calls Claude is separated from the sandbox where the agent's code runs; the agent loop, session history and vault secrets sit outside the sandbox; each environment can be locked to a list of allowed hosts, with requests to anything else blocked; and every step an agent takes is logged and exportable over OpenTelemetry. Credentials stay out of the sandbox too, and self-hosted customers can keep them in their own secrets manager. Organisations with a zero data retention agreement with Anthropic have those model requests covered the same way as any other call to the Messages API.

Illustrative image: a data centre. Self-hosted Managed Agents runs on the customer's own cloud or on-premises infrastructure — the photo shows a data centre, not Anthropic's service. Image: BalticServers, Wikimedia Commons.

What Is Missing in the Private Beta

The private beta is not a feature-for-feature copy of the hosted service. Scheduled deployments, dreams and webhooks are unavailable on the self-hosted server, and web search and web fetch stay switched off until a team explicitly enables them. Anthropic's documentation describes a Docker Compose setup for development and warns it is not appropriate for production, pointing teams toward Helm-based deployment on a Kubernetes cluster. There is also no migration path: agents, skills, vault credentials and memory stores built on the hosted service have to be recreated. And Anthropic has published no pricing, no list of covered countries and no date for general availability, so the beta is for eligible organisations that already have an Anthropic account team, not something a small team can sign up for and try this weekend.

Setup	Who runs the agent server	Where credentials and history live	Model requests go to
Hosted service	Anthropic	Anthropic	Anthropic (or chosen cloud provider)
Self-hosted sandboxes (May 2026)	Anthropic	Anthropic — only code execution moves home	Anthropic (or chosen cloud provider)
Fully self-hosted (private beta, 12 Nov 2026)	Your company	Your cloud or data centre	Provider you pick: Anthropic, Bedrock, Google, Foundry or your own gateway

Why This Matters for Readers in India

Most readers will never run a self-hosted agent server — this is enterprise infrastructure, not a consumer feature. But the timing lands on a live Indian question. India's DPDP Rules opened consent-manager registration on 13 November 2026, and banks, healthcare providers and government departments are all being asked where their data sits and who can touch it. For those organisations, an AI agent that holds production credentials and a full session history is a data-governance problem — and keeping both inside their own cloud is the part of this announcement that actually answers it. The honest limit matters too: because prompts still travel to the chosen inference provider, self-hosting is not full data localisation, and Anthropic has published no India-specific availability for the beta. For Indian developers, the nearer value is the pattern: the same week Google put Gemini 4 Hydrogen into Search, Anthropic pushed Claude further into office work with Claude Sheets and Whiteboards — the frontier labs are racing on the plumbing around the model, and that plumbing is where enterprise adoption is decided.

Should you care?

✅ Yes, if your team already trusts Claude agents with real systems — self-hosting removes the biggest objection to letting agents work unattended inside a company: credentials and history no longer live on someone else's infrastructure. But it trades that for operational work — patching, securing and running the server yourself — and the model traffic still leaves your network. If you only use Claude as a chatbot, this changes nothing for you yet.

FAQs

What is Claude Managed Agents?

Claude Managed Agents is Anthropic's hosted service for running Claude as a long-running autonomous agent. You define an agent once — its model, instructions, tools and skills — then start sessions that work in their own sandboxes, running code, editing files and searching the web. It launched as a hosted beta in April 2026.

What is the difference between self-hosted sandboxes and self-hosted Managed Agents?

With self-hosted sandboxes, added in May 2026, only the isolated environment where the agent's code runs sits on your infrastructure — Anthropic still runs the service, including the API, the agent loop and the session history. With fully self-hosted Managed Agents, you run the whole service: the server, the database holding session history, and the workers, with only the model calls going to an outside provider.

How do I get access to self-hosted Claude Managed Agents?

It is a private beta, and the only channel is Anthropic's account team, which hands over the release archive for installation. There is no public signup page, no published pricing and no general-availability date yet.

Does self-hosting Claude Managed Agents keep my data inside my network?

Only part of it. Credentials, session history and the sandboxes stay on your infrastructure, but every request to Claude — prompts, tool inputs and tool results — still travels to the inference provider you choose.

Takeaway: Anthropic has handed enterprises the keys to the agent platform — the control plane, the credentials and the history now live behind the customer's own perimeter. The model is still a guest, not a resident, and the beta's missing features and unpublished pricing mean it is a starting point for regulated teams, not a finished product.

If your team already runs Managed Agents, the useful next step is to read Anthropic's self-hosting documentation and ask your account team what the beta covers — before assuming it solves a compliance requirement on its own.

Sources

• Anthropic — Claude Managed Agents documentation (overview)
• Help Net Security — "Self-hosted Claude Managed Agents keeps agent credentials in your environment" (12 November 2026)
• AI Frontier Post — "Anthropic lets enterprises self-host the full Claude Managed Agents stack — private beta now open"
• Derrington Collaborative AI — "You can now host Claude Managed Agents yourself"