What Anthropic shipped on 12 November
Anthropic's release notes for 12 November 2026 add a third way to run Claude Managed Agents. In the default setup, Anthropic hosts everything. Since 19 May 2026, self-hosted sandboxes have let teams run the tool side themselves, while Anthropic kept the API, the agent loop and the stored sessions. With the new self-hosted option, you take that second half as well. Agents, sessions and each session's full history go into a database you run, either in your own cloud account or in an on-premise data centre.
You can't sign up on a web page. Both the download and the pricing go through your Anthropic account team. The Next Web says the launch comes as European governments and companies worry more about data residency and about relying on foreign AI vendors. It also points out that the model itself still runs outside your network.
The moving parts
A deployment is one or more servers, a database and your sandboxes. Here's what each piece does:
- cma-server takes API requests, runs the agent loop, keeps all the records and calls the model when the agent needs it. It never runs the agent's shell or file tools itself.
- PostgreSQL 16 or newer holds the data. Without it, the server keeps everything in memory and loses it all on a restart.
- Workers poll the server for jobs and run tools inside sandboxes. You start one with
ant beta:workeror with an SDK. - A web console sits on the server's own address. People sign in to it through your OpenID Connect identity provider.
- cma-server-admin is a CLI for creating environments and their keys. It runs on Linux (x86-64 and Arm) and on Apple silicon Macs.
You'll handle three secrets. Your apps and CLIs use the server API key to talk to your server, the same way they'd use a Claude API key with the hosted service. Only the server holds the inference credential, which pays for model calls. Workers use an environment key, which lets them pull work from one environment's queue. Your server exposes the same API as the hosted service, so the SDK and the ant CLI only need a new base URL.
Running the quickstart
Before you start, you'll need:
- the release archive
- a Claude API key
- a current Docker Engine with the Compose plugin
openssland theantCLI- a machine with at least two CPUs. Smaller hosts should set
CMA_SANDBOX_CPUS.
The first build also needs internet access, because it pulls from Docker Hub, the Alpine package repositories and GitHub releases.
Unpack the archive and work from its compose folder. If your umask is stricter than 022, the server container can't execute the binaries, so reset it and extract again. Next, write a .env file with four values: a Postgres password, your Claude key, a server API key and a key that encrypts stored secrets. The quickstart uses openssl rand to generate the three random ones, then locks the file down with chmod 600.
Run docker compose up --build -d and watch the cma-server log until it says it's listening. Then call the readiness endpoint at http://127.0.0.1:8400/cma-server/readyz. It's normal to see "degraded" for a few seconds before "ready" appears. If you see "not_ready", the response names the check that failed, usually the upstream key or the database.
Open a second terminal for ant. Point ANTHROPIC_BASE_URL at the local server, and set ANTHROPIC_API_KEY to the server key from .env, leaving your Claude key alone. Keep Compose in the first terminal. If you start the stack from the shell that holds those variables, the server picks them up and ends up sending model calls to itself.
You define the agent in a small markdown file. Its front matter gives a name, a model (claude-opus-5-5) and the built-in toolset agent_toolset_20260401. Load it with ant apply --yes. You'll see a 404 warning about the organisation and workspace, which is expected on a self-hosted server. The stack comes with an environment called compose-sandbox, so you can start a session straight away. Ask the agent to run hostname. It'll answer with a name starting cma-sandbox-, which shows the command ran in a container on the stack and not on your laptop. docker ps --filter label=cma-server.sandbox lists that container. It shuts down about a minute after each turn.
The stack runs 10 workers, and each one serves one session at a time. A session that never gets a message can hold a worker for up to an hour. If tool calls start hanging, archive the sessions you aren't using. docker compose down --remove-orphans stops everything and keeps the database. Add -v if you want to wipe it.
Treat this stack as a dev toy. It serves plain HTTP. It binds port 8400 to loopback, but some Docker Engine releases before 28.3.3 still let other machines on the LAN reach it. Use a spare machine with nothing valuable on it. For production, Anthropic points you to two Helm charts for Kubernetes. The optional console demo uses a mock sign-in provider on port 8080, and that provider gives an admin token to anyone who can reach it.
Choosing where model calls go
The inference.provider setting accepts five values: anthropic (the default), bedrock, vertex, foundry and gateway, which covers a model gateway you run yourself. Agents keep using Claude API model IDs, and the server translates them for each provider.
Your choice changes which features work:
web_searchworks underanthropic,vertexandfoundry.web_fetchworks only underanthropicandfoundry.- Both web tools stay off until you turn them on in the server's settings file.
inference_geois only accepted underanthropic.- Fast mode depends on both the provider and the model.
What you give up
Nothing carries over from the hosted service, so you'll rebuild agents, skills, vault credentials and memory stores on your server. Several features are missing entirely:
- cloud environments
- the Messages and Models APIs, which your server doesn't serve
- the Files API. Put files in the worker's directory, and send images and PDFs as base64.
mcp_oauthcredentials- the prebuilt docx, pdf, pptx and xlsx skills, though you can upload their sources from Anthropic's skills repository
- scheduled deployments, dreams and webhooks
The server only connects to MCP and Git hosts you've put on an allowlist. By default it cuts MCP results at 100 KiB.
Security is now your job. The agent's bash tool is a real shell, with the same permissions and network access as the worker process. The only isolation is whatever container or VM the worker runs in. Anthropic advises a separate deployment for each tenant or trust level, because a single server API key reaches every environment.
Costs
For self-hosting, Anthropic only says to ask your account team. On the hosted service, sessions pay token rates plus $0.08 USD for each hour a session spends running. In Anthropic's worked example, a one-hour Opus 5.5 session with 50,000 input tokens and 15,000 output tokens costs $0.58 USD. If you self-host, add your own servers and Postgres. Model usage falls under your agreement with whichever provider serves the model.
What this means for New Zealand builders
You decide where transcripts are stored, which could be a data centre here. The server keeps sessions until you delete them, so you'll need your own retention policy. The prompts and tool results in each model call still go to your inference provider. Calls to the Claude API also carry session, agent and environment ID headers, and no setting turns them off.
Anthropic's data retention page lists hosted Managed Agents as outside zero data retention, because Anthropic stores the transcripts. When you self-host, you store them, and an existing ZDR arrangement covers the model calls just like other Messages API traffic.
We think a good first project is an internal coding agent with a Git allowlist that only lets it push to refs/heads/agent/ branches. Another option is a pilot for a regulated team that needs every tool call logged on its own systems.
Sources
- Claude Platform release notes
- Self-hosted Claude Managed Agents
- Get started with self-hosted Claude Managed Agents
- Inference provider
- Differences from the hosted service
- Data and access in self-hosted Claude Managed Agents
- Pricing
- Self-hosted sandboxes
- The Next Web: Anthropic self-hosted Claude Managed Agents
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.