Image: Anthropic
Why it mattersA team can now run the agent service inside its own network. The licence makes the team roll back a flagged version within eight hours.
A team can now run the Claude agent service inside its own network. The catch is in the licence.
Anthropic opened self-hosted Claude Managed Agents in private beta. The announcement came from Anthropic's ClaudeDevs account. The company has published a docs page and a licence, and access is through a team's Anthropic account manager.
What the team hosts
The documentation lists three options. The default is cloud sandboxes, where Anthropic runs the service, the sandboxes and the model. The second is self-hosted sandboxes, where the team runs only the sandboxes. The new third option is self-hosted Managed Agents, where the team runs the service too.
The service is one or more servers, a PostgreSQL 16 database and a set of workers. The workers run the agent's shell and file tools in the team's sandboxes. Anthropic says the server stores every agent, session and session history. Workers poll the server for work. The model runs at Anthropic by default. It can also run at Amazon Bedrock, Google Cloud, Microsoft Foundry, or the team's own model gateway.
The eight-hour clause
The licence sets a hard rule for safety. Section 5.3 covers what happens when Anthropic flags a safety, security or misuse risk in writing. The team must disable that feature or roll back that version within 8 hours. The rule applies to every deployment the team runs.
A normal update has a 30-day window in the same document. The 8-hour clause is for safety and security only. A team running the service across many customers has to be ready to act fast across every deployment.
What leaves the team's network
The server sends the full conversation to the team's inference provider for each model call. That is the one piece of data that goes out. By default the call goes to the Claude API. It carries the session ID and the agent and environment IDs with it. A team that routes the model to its own gateway can keep even that call inside its network. The agent's shell and file tools run in the team's sandboxes.
Three credentials, not one
The documentation names three keys. A server API key lets the team's applications and CLI talk to the server. An inference credential lets the server call the chosen model. An environment key lets workers poll a queue of work. The server never runs an agent's shell.
Anthropic warns that the bash tool is a real shell with the worker's permissions and network access. The worker itself does not isolate it. The team has to run each worker in a disposable container or VM. Anthropic provides a Docker Compose stack and a Helm chart that do this for each session.
Source
- Self-hosted Claude Managed Agents, Anthropic documentation.
- Software License Agreement: Claude Managed Agents Server, Anthropic.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
