Claude Agents Can Now Run Entirely Inside Your Own Walls

Human Verified 0

Share to: , and

A steel security door beside a glass-walled data centre

Anthropic has opened a private beta that lets companies run Claude’s agent service entirely inside their own systems, the company announced on Thursday. Self-hosted Claude Managed Agents moves the whole service — the agent harness, sandboxes, credentials and session history — into a customer’s own cloud or data centre. It is the furthest Anthropic has gone toward letting enterprises keep autonomous agents behind their own security perimeter.

What customers now run

Managed Agents is Anthropic’s suite of composable APIs for building and deploying autonomous agents, covering the harness, sandboxes, credential handling and session history that get long-running agents into production. Until now it ran as a hosted service on the Claude Platform, where Anthropic operated the infrastructure. Self-hosting moves every part of it into the customer’s environment.

The customer operates three things: the Managed Agents server, its database, and the workers where agents run their shell and file tools. The server exposes the same Managed Agents API as the hosted service, apart from documented differences, so the Claude SDK and the ant command-line tool can connect to it. A persistent deployment needs PostgreSQL 16 or later.

A developer's desk with two monitors showing flowing lines of light

Model requests go where the customer chooses. The default is the Claude API, but teams can route calls through Amazon Bedrock, Claude Models on Google Cloud, Microsoft Foundry, or a gateway they operate themselves. Session history stays in the customer’s own database, and zero data retention applies to Claude requests as it does to any Messages API call.

Built for unattended agents

Anthropic built Managed Agents around four security principles, which apply whether it hosts everything or the customer hosts the whole service. The harness and the sandbox are split by design, so an agent’s code runs apart from the agent loop, session history and vault secrets. Credentials stay out of the sandbox: a proxy outside it adds them to requests.

Network access can be limited to a list of allowed hosts, with other requests blocked. For an agent that fixes broken builds, the list might be just the git host and the package registry. Every session leaves a full record of each step, including the tool calls made and what came back, exported over OpenTelemetry to the team’s monitoring tools.

Our Read

Read plainly, this is Anthropic answering the question every enterprise asks about agents that work while nobody watches: who holds the keys? Moving the whole service behind the customer’s perimeter removes the biggest objection to letting agents touch production code. The trade is real, though: the customer operates the server, the database and the workers, and installs each release itself.

The caveat is that self-hosting moves the control plane, not necessarily the model calls. By default those still go to Anthropic, and only a customer’s gateway keeps them inside its network. Migration is a fresh start: nothing is copied from the hosted service, so agents, skills and credentials must be recreated. Access is a private beta, through an account team.

Why It Matters

It matters because agents that work unattended are the next enterprise battleground, and trust is the bottleneck. Anthropic’s pitch is that the same agent service can run inside a company’s walls, with the same APIs and security principles. If that holds, compliance teams lose their main objection — and rivals must match the deployment choice, not just the model.

The timing is part of the story. Anthropic launched the option the same day it took Claude into manufacturing plants, and a week after Claude grew an office suite — a coordinated push to make Claude the place where work happens. Mistral’s Studio already turns its platform into an agent factory, so the self-hosting option is a competitive answer.

Background

Managed Agents began as a hosted service on the Claude Platform, and Anthropic later added self-hosted sandboxes for teams whose code had to stay in their own environment. The new option extends that choice to the whole service, including the harness and session history. Anthropic has not said when the private beta will open more widely.

Source: Anthropic’s announcement and Help Net Security’s report.

Share to: , and

Faheem R. Khattak

Senior automation engineer and AI marketing journo covering LLMs, algorithms, analytics, core updates, Bing search, and technical SEO.

0 Comments

Leave a Reply

Optional Take Tag:

🛡️ Spam protection: Honeypot + Time-Trap Active

This site uses Akismet to reduce spam. Learn how your comment data is processed.