Anthropic lets companies run the whole Managed Agents service themselves

Anthropic has launched self-hosted Claude Managed Agents, according to its Claude Platform release notes dated November 12, 2026. Customers can now run the entire Managed Agents service, and store agents, sessions and session history, on their own infrastructure. Access is by request only.

The service moves in-house, not just the sandboxes

Until now, Anthropic offered two arrangements for Managed Agents: its own cloud sandboxes, or self-hosted sandboxes where only the tool execution ran at the customer. The new option goes further. According to the release notes, the customer runs the whole service, and the server sends model requests to a provider of the customer’s choice, with Anthropic as the default.

The documentation describes the split in three parts. The server, called cma-server, serves the API, runs the agent loop and stores agents, sessions and history. A PostgreSQL database, version 16 or later, holds the data; without one, the server keeps everything in memory and a restart erases it. Workers, started with ant beta:worker or an SDK worker, run the agent’s shell and file tools in the customer’s sandboxes. The server itself never runs those tools.

Anthropic says the server exposes the same API as the hosted service, so the SDK and the ant CLI work with it. The documentation also describes a web console on the server, where people sign in through the company’s own identity provider, and an admin CLI, cma-server-admin, that creates environments and their keys.

Who can use it, and what it demands

The release notes say to contact an Anthropic account team to request access, and the documentation repeats that for the download. No public pricing, general availability date or beta label appears in the material Anthropic published. The deployment can sit in a cloud account or an on-premise data center.

The documentation names two reasons to choose it: storing agents, sessions and history on your own infrastructure, and sending model calls somewhere other than Anthropic. Listed alternatives are Amazon Bedrock, Claude Models on Google Cloud (Google managed), Microsoft Foundry or a company’s own model gateway. Companies that only need tools to run on their hardware are pointed to the existing self-hosted sandboxes instead.

What the source leaves open

The trade-off is operational. Anthropic’s documentation says the customer runs and secures the server, its database and the sandboxes, and installs each release by hand. It warns that the agent’s bash tool is a real shell with the worker’s permissions and network access, and that isolation comes from the container or virtual machine the worker runs in. What Anthropic’s security model assigns to itself on the hosted service, including securing the server, hosting memory stores and spotting unusual usage, becomes the customer’s job.

Data does not fully stay home either. For each model call the server sends the conversation so far, including tool inputs and results, to the configured inference provider. Calls to the Claude API also carry the IDs of the session, agent and environment. Anthropic adds that some features exist only on the hosted service and points to a separate page of differences. The release notes do not list them, and this article could not verify which are missing.

Multitenancy is also strict: the documentation says each trust level needs its own deployment. Companies serving several internal groups with different data rules should expect to run more than one server.

The move extends a path Anthropic started earlier this year. Secondary coverage, such as The Decoder, reported that self-hosted sandboxes arrived on May 19, 2026 in public beta, with orchestration still at Anthropic. Moving orchestration and storage too answers the objection that session data had to live on Anthropic’s side.

The timing is notable. The announcement lands days after Anthropic launched the roadshow for its proposed public offering, which we covered in our roadshow report, and three days after it released Claude Fable 5.2, which Copilot admins must accept data retention terms to use. Anthropic does not link the releases, and the source says nothing about customer demand or how many organizations have access.