Self-host AI agents: what the private beta means for SMEs
On November 12, 2026, Anthropic opened a private beta that lets companies self-host AI agents. Self-hosted Claude Managed Agents places the agent service in your own cloud or on your own servers: agents, credentials and session history stay there. The request to the model goes to the provider you choose. The default is Anthropic.
For an SME this is not a switch to flip on Monday. You request access through your Anthropic account team, and you operate and update the server, the database and the workspaces yourself. The beta still answers a question that comes up before the first production agent: what should sit on infrastructure you operate, and what leaves with a model call? How an AI agent works in general is on our page about AI agents.
This article reads the announcement and the Claude documentation as of November 15, 2026. Lyron has not tested the private beta. Every technical detail comes from Anthropic.
In short
- Private beta since November 12, 2026: You run the whole Managed Agents service on your own infrastructure, including the sandbox. Access goes through the account team.
- On your side: agents, sessions, session history, credentials and the environment where the tools run.
- Still sent out: By default, each model request sends the conversation and the tool results to Anthropic, unless you choose another provider. The model runs at that provider.
- Nothing is copied over: Agents, history and credentials from Anthropic’s hosted service are created again on your server.
- For most SMEs: A narrow workflow in n8n is still the closer first step. The beta matters when the agent has to work inside internal systems and someone will operate it.
What opened on November 12
Anthropic describes Managed Agents as a set of interfaces for long-running agents: the harness, isolated workspaces called sandboxes, credential handling and session history. It first shipped as a service on the Claude Platform, with Anthropic running the infrastructure. Self-hosted sandboxes came later, for companies whose code had to stay in their own environment.
On November 12, 2026, the Claude Platform release notes add the rest. They say you run the whole service on your own infrastructure and store agents, sessions and session history there. The server sends model requests to the provider you choose. The default is Anthropic. You request access from your account team. The announcement describes the same private beta and the same scope: your own cloud or on-premises, including credentials and session history.
Three ways to run the agent
The self-hosted documentation separates two parts. The sandbox is where tools run commands and work on files. The service is the API, the agent loop and the storage. That produces three setups:
| Setup | Sandbox | Service | Model |
|---|---|---|---|
| Cloud sandbox, the default | Anthropic | Anthropic | Anthropic |
| Self-hosted sandbox | You | Anthropic | Anthropic |
| Self-hosted Managed Agents | You | You | Anthropic, or a provider you choose |
The middle row is enough when files and processes should stay on your machines. The page Choosing an agent environment says that whatever a tool returns to the model still reaches the service Anthropic operates. Web search and web fetch run on Anthropic’s side, outside the sandbox.
The third row is the November 12 beta. The documentation names two reasons to choose it. Agents, sessions and session history have to live on your infrastructure, or model calls have to go to a provider you name, such as Amazon Bedrock, Claude models on Google Cloud or your own gateway.
What runs on your infrastructure
The announcement describes four properties that matter in production. They apply whether Anthropic hosts everything, you host only the sandbox, or you host the whole service.
- The harness and the tools are split. The harness calls the model. The sandbox runs the code. The agent loop, the session history and stored secrets stay outside the sandbox. A sandbox starts when a tool needs it.
- Credentials stay outside the sandbox. For MCP tools and supported API calls, a proxy adds the credential to the request. On the self-hosted service you can keep secrets in your own secrets manager.
- Network access can be limited. Each environment can be restricted to a list of allowed hosts. Requests to any other host are blocked. The announcement uses an agent that fixes broken builds and may reach only the git host and the package registry. When you host the sandbox, you set the same limits in your firewall.
- Every session leaves a record. It includes the tool calls and what came back. The record exports over OpenTelemetry, so you can review the activity in the monitoring you already use. How runtimes and stops show up day to day is covered in monitoring AI workflows.
What each model call still sends out
Running the service yourself means the session history sits in your database. It also means the model runs wherever you send the request. The environment guide states the default plainly: each model request still sends the conversation history and the tool results to Anthropic. The model runs at Anthropic or at a provider you choose.
If your organization has a zero data retention agreement with Anthropic, the announcement says it applies to these requests the same way it applies to any other Messages API call. That describes retention at Anthropic. With the default provider, the request still leaves your network.
Whether that split fits your contracts, your processing terms and the data inside the work is a decision for your company. This article is practical guidance and does not replace legal advice.
Example scenario from an SME
Example scenario, not a customer case: A mechanical-engineering supplier in North Rhine-Westphalia with 40 employees keeps order notes in an internal system. Management wants an AI agent to draft a status update for sales from those notes. The session history should stay on a server the company controls.
In the beta that would be the third setup: service and sandbox on the company’s network, model requests to the chosen provider. In the scenario the company does not take the beta as the first step. Nobody there operates an extra agent server today, and access is a private beta. Instead, a workflow in n8n pulls the open orders, lets the model write only the draft and puts the text in front of a person before it goes anywhere. The team keeps the self-hosted service in mind for later, if the agent has to work inside the internal systems and someone will operate it. What that approval looks like is in the guide to permissions and approvals. How tight the instruction should be is in the guide to instructing AI agents reliably.
What you would operate yourself
The documentation lists the parts that then sit with you: a server for the API and the agent loop, a database, and workers that run commands and file access inside the sandbox. The named database is PostgreSQL 16 or later. Without a database the server keeps its data in memory, and a restart erases it. You install each release yourself.
An agent’s shell is a real shell with the worker’s permissions. Each session runs in its own container or virtual machine. The documentation points to the Docker Compose stack and the Helm chart, which start that worker separately for each session.
The page Differences from the hosted service names limits that matter before a trial. Nothing is copied from the hosted service, session history included. You create agents, skills, stored credentials and memory stores again. Scheduled runs and webhooks are absent on your server. You start sessions from your own scheduler and follow a session through its event stream. The server rejects prebuilt document skills. Only skills you upload are available.
The same API, fewer ready-made pieces, and the operations sit with you. That is the cost of running the service yourself.
When an n8n workflow is the closer fit
Most work in a small or mid-sized company does not need an agent working unattended inside internal systems. It needs a workflow someone can read back: fetch the data, let the model handle one bounded step, check the result, write it onward. n8n can be self-hosted for that. Where the data then lives, and what you update yourself, is covered in n8n self-hosting versus cloud. The wider frame is on the page about process automation.
An AI agent earns its place when the path changes from case to case and fixed steps are not enough. Limits and a human approval still belong at the start. The private beta is the relevant look when three things come together: the agent has to work inside your internal systems, history and credentials have to sit on infrastructure you operate, and someone owns the server, the database, the updates and the network limits. If one of those is missing, a narrow workflow is the more practical first step.
What to check before a trial
- Confirm access: It is a private beta. The request goes to the account team. The documentation does not describe a public switch.
- Write down the data flow: What stays on your server, and what leaves with each model request.
- Keep permissions small: The shell has the worker’s permissions. A separate environment per session, an allowlist of hosts, and approval before anything is written.
- Plan to recreate: An agent from the hosted service is not brought across. You create it again on your server.
- Name an operator: Updates, PostgreSQL, workers and the session record need a person responsible for them.
- Start with one narrow case: One job, one draft, one person approves. Widen it once the workflow holds up in daily use.
Sources
As of November 15, 2026. Technical details come from Anthropic. The reading and the example scenario are Lyron’s.
- Claude Platform release notes – November 12, 2026 entry on the private beta.
- Anthropic: Operate Claude Managed Agents inside your security perimeter – scope of the beta, split between harness and sandbox, credentials, network limits, session record.
- Self-hosted Claude Managed Agents – the three setups, the database, workers and where the model runs.
- Choosing an agent environment – what a self-hosted sandbox still sends to Anthropic, and that model requests send the conversation and tool results by default.
- Differences from the hosted service – nothing is copied; scheduled runs, webhooks and prebuilt document skills are absent.
Want an AI agent where it is clear where history and credentials sit?
Lyron sets up a narrow first workflow with you, for example in n8n: the model writes a draft, a person approves it, and your team can maintain the workflow.
Book a free consultation