What Anthropic released
On 12 November, Anthropic announced a self-hosted version of Claude Managed Agents, its service for building and running long-running autonomous agents. The release is in private beta, and organisations request access through their Anthropic account team. Before this release, customers could only move the sandbox onto their own infrastructure. The sandbox is the workspace where an agent's tools run commands and edit files. When Anthropic launched self-hosted sandboxes in May, the agent loop and context management stayed on Anthropic's servers, and there was no fully on-premise option.
The new option moves the rest of the service across. Anthropic's documentation describes a customer-run server that serves the API, runs the agent loop and stores agents, sessions and session history in a PostgreSQL 16 or later database. Worker programs poll that server and run the agent's shell and file tools inside the customer's sandboxes. The model still runs at Anthropic or at a provider the customer chooses: Amazon Bedrock, Claude models on Google Cloud, Microsoft Foundry or the customer's own model gateway. VKTR reports that Anthropic didn't disclose pricing.
What still leaves your network
Moving the agent service in-house reduces how much leaves your environment. Model calls still carry a lot. According to the documentation, each model call sends the conversation so far to the inference provider, and that includes every tool input and tool result. Suppose an agent pulls a customer record from your CRM through an MCP tool. That record goes to the model provider on the agent's next turn. Calls to the Claude API also include the session, agent and environment IDs.
For NZ organisations, the choice of inference provider decides where that conversation is processed. On 26 March 2026, AWS made Bedrock available in its Auckland region, with Claude models offered through an Australian cross-region profile. Requests that start in Auckland are either served locally or routed to Sydney or Melbourne. Data at rest stays in the source region. A self-hosted server in AWS Auckland that sends model calls to Bedrock would keep sessions, credentials and history in New Zealand, and model processing would stay within New Zealand and Australia. When personal information goes to an overseas provider, firms should check how information privacy principle 12 applies. IPP 12 sets the conditions for disclosing personal information outside New Zealand.
Some features depend on which provider you choose. The list of differences from the hosted service says the inference_geo setting only works with Anthropic's own API. Built-in web search isn't offered when calls go through Bedrock. When the Anthropic API is the provider, the server turns on server-side fallback by default. If a safety classifier declines a request, a different and possibly less capable Claude model can answer it, and no event records the switch. Teams that need a record of which model produced each step can turn fallback off with a single setting.
What you take on
Most of the cost is operational. The documentation says self-hosted customers run and secure the server, its database and the sandboxes, and they install each new release themselves. Security duties that Anthropic carries on the hosted service move to the customer, including detecting unusual usage and hosting memory stores. The documentation also warns that the agent's bash tool is a real shell with the worker's permissions and network access. The worker program doesn't isolate that shell, so isolation depends on the container or virtual machine the worker runs in. Each trust level needs its own deployment. A firm with a finance agent and a customer-facing agent that reach different data should plan for two servers.
Several hosted features aren't available. The self-hosted server sends no webhooks and doesn't support scheduled deployments, so recurring jobs need your own scheduler. The server rejects Anthropic's prebuilt skills for Word, PDF, PowerPoint and Excel files, so teams have to upload them as custom skills. It also rejects MCP credentials that refresh through OAuth. Teams have to store those tokens as static bearer credentials and replace them by hand before they expire. Nothing carries over from the hosted service, including session history.
Some of the defaults are strict, which helps security. MCP servers must be on an allowlist that starts empty, web tools stay off until an administrator enables them, and Git pushes are limited to branches the administrator lists. The server doesn't refuse force pushes, so branch protection on your Git host has to stop an agent from overwriting a listed branch. Anthropic's announcement says every step is recorded, including tool calls and their results, and can be exported through OpenTelemetry into existing monitoring tools.
What this means for NZ businesses
We expect most mid-sized NZ organisations to start with the hosted service or with self-hosted sandboxes. Full self-hosting suits organisations that must hold agent session history themselves, such as health providers, financial services firms and government suppliers, or whose contracts specify the inference route. Session history is the agent's working record. It holds every prompt, tool result and intermediate draft, and it often contains more sensitive detail than the finished output. After an incident, that record is what investigators and auditors will ask to see. Keeping it on your own infrastructure puts the record under your control.
Self-hosting also gives your own team the patching, monitoring and incident response for a platform that runs real shell commands. Organisations that already fall behind on server patching would add risk by taking on an agent platform with a shell inside it. For those teams, the hosted service with tightly scoped tools is the safer configuration. Our AI agents page covers how we scope agent access. For firms weighing how much of their AI stack to run in-house, our private AI page covers the infrastructure side.
Practical next steps
- List what each planned agent would read through its tools. Every tool result goes to the model provider on the next turn, so the tools set your real data exposure.
- Pick the inference route before you build. If processing needs to stay in Australia and New Zealand, look at Bedrock in Auckland with the Australian profile, and confirm the models you need are in that profile.
- Compare your design with Anthropic's list of differences. Missing webhooks, scheduling and document skills, plus OAuth tokens that need manual rotation, can each force a redesign.
- Run every session's worker in a disposable container or pod. Set up one deployment per trust level, and enforce host allowlists in your own firewall.
- Name owners for release upgrades, token rotation and the OpenTelemetry feed into your monitoring, before the first agent goes live.
- Turn off model fallback if you need to trace which model produced each step. Protect agent branches on your Git host.
- Ask your Anthropic account team about access and pricing before you budget. Neither has been published.
Sources
- Operate Claude Managed Agents inside your security perimeter (Anthropic)
- Self-hosted Claude Managed Agents (Claude Platform Docs)
- Differences from the hosted service (Claude Platform Docs)
- Anthropic moves Claude Managed Agents inside the enterprise perimeter (VKTR)
- Anthropic adds self-hosted sandboxes and MCP tunnels to Claude Managed Agents (The Decoder)
- Run generative AI inference with Amazon Bedrock in Asia Pacific (New Zealand) (AWS)
- Privacy principle 12 (Office of the Privacy Commissioner)
- AI Agents News, week of 17 November 2026 (AI Agent Store)
Want to know what this means for your organisation? Our team in Auckland can show you your top three opportunities in a free 30 minute assessment.
Get my free assessment How we help