- Anthropic launched self-hosted Claude Managed Agents on 12 November. You run the API, the agent loop, the database and the sandboxes. Model calls go to the Claude API by default, or to Bedrock, Google Cloud, Microsoft Foundry or your own gateway. Access goes through an Anthropic account team.
- The server exposes the same API as the hosted service, with a page of exceptions. In seven of them nothing errors: an image sent by URL never reaches the model, a session with no worker stalls at its first tool call, and a push your Git host refused is logged as
ok. - With the Claude API as provider, every model call carries three headers with your session, agent and environment IDs, and the docs say there is no setting to turn them off. Without a database, which is the default, a restart erases every agent and session.
Since May you could run the sandboxes of Claude Managed Agents on your own machines while Anthropic ran everything else: the API, the agent loop and the session history. The release notes for 12 November add the rest. You can now host the service itself and keep agents, sessions and their history in your own database.
I haven't run it. The download goes through an Anthropic account team and I don't have the archive. What follows is a read of the launch documentation, mostly two pages: the one listing differences from the hosted service, and the one on data and access. I'd read both before the quickstart.
What you run
A deployment has three parts. A server called cma-server serves the API and runs the agent loop. A PostgreSQL database, version 16 or later, stores agents, sessions and history. Workers poll the server for tool calls and run them in your sandboxes. The server never runs the agent's shell or file tools itself.
Two defaults matter on day one. If you configure no database, the server keeps everything in memory, and a restart wipes it. And the Docker Compose stack in the quickstart is marked as development only. It serves plain HTTP, and Anthropic points production installs to two Helm charts.
Why take this on
Anthropic names two reasons. Either you must store agents, sessions and history yourself, or your model calls have to go to Amazon Bedrock, Google Cloud, Microsoft Foundry or a gateway you operate. If you only need the tools to run inside your network, the docs send you back to self-hosted sandboxes.
The storage reason has a compliance side. Hosted Managed Agents is listed as outside zero data retention, because Anthropic stores the session transcripts until you delete them. On your own server you store them, and the docs say a ZDR arrangement then covers the model calls like any other Messages API call. For a regulated team, that alone can justify the extra operations work.
Seven cases with no error
The pitch is one API for both, so the SDK and the ant CLI work against your server unchanged. That is also the risk. Code written for the hosted service keeps getting successful responses where the feature behind them is missing.
I counted seven where the docs describe no error at all.
| What you do | What your server does |
|---|---|
Send an image or document block with a url source |
Accepts it. The model never receives the content. An image with a file source goes the same way. |
| Start a session when no worker is polling its environment | Runs the model turns. The first tool call then waits until a worker turns up. |
| Run an agent whose toolset doesn't mention the web tools | web_search and web_fetch are off by default, so the agent runs without them. |
List web_search on Bedrock or a gateway, or web_fetch anywhere except the Claude API and Foundry |
Saves and runs the agent with the tool still listed. The model is never offered it. |
| Let an agent push to a branch your Git host rejects | Logs ok for the push. |
| Start your workers from a webhook | Sends no webhooks, so those workers never start. |
Filter sessions on the rescheduling status |
Retries failed model calls but keeps the status at running, so the filter returns nothing. |
The first row is the one I'd expect to bite first, since passing an image by URL is the default in a lot of client code.
Three more fail late or only in part. A custom tool result too large for the context window gets a 200, and then the session ends with a session.error. An MCP tool result over 100 KiB is cut at that size. The model is told, and the rest is dropped, where the hosted service can save the full result to a file.
The third is model fallback. With the Claude API as provider, the server asks for server-side fallback by default. When a safety classifier declines a request, another Claude model, possibly a weaker one, can answer instead. No event reports the swap. Reading the session shows the model ID, and inference.model_fallback: false turns it off.
The fix for the image case is to send the bytes as base64. The others have no request-side fix. They need a check you write yourself: a worker heartbeat, a look at the branch after a push, a test that the agent can search.
What isn't there
Some hosted features are gone, and those do return errors. A request for a cloud environment gets a 400. The Messages API and the Models API aren't served, so SDK calls to /v1/messages come back 404. There is no Files API, no scheduled deployments, no dreams and no webhooks.
mcp_oauth vault credentials are rejected, because the server has no OAuth flow to refresh a token. You store a static bearer token and rotate it before it expires. The prebuilt docx, pdf, pptx and xlsx skills are missing until you upload their sources as custom skills.
Multiagent orchestration shrinks to a coordinator with 1 to 20 subagents that you list. The advisor, dynamic workflows and subagents the agent defines itself are out. The docs add that their own multiagent examples, which set type to multiagent_20261001, return a 400 on your server.
Nothing is copied over from the hosted service, session history included. You create your agents, skills, vault credentials and memory stores again.
In return the server has four things the hosted service lacks: the choice of inference provider, single sign-on through your OpenID Connect provider, vault secrets kept in your own secret manager, and agent identity, a short-lived signed token sent along with an agent's calls to MCP servers.
What still reaches Anthropic
The server connects to Anthropic only when the Claude API is its inference provider. On Bedrock, Google Cloud or Foundry it makes no connection to Anthropic, and Anthropic has no remote access to the server or its log.
With the default provider, each model call carries the conversation so far, including tool inputs and results. It also carries three headers holding the session ID, the agent ID and the environment ID your server generated. An ID contains no name and no content, but it does encode when it was created, and Anthropic can use the headers to attribute usage. The data page deals with opting out in five words: "No setting turns them off." If ANTHROPIC_BASE_URL points at a proxy of yours, the headers go to that proxy.
Three smaller things also reach Anthropic. A User-Agent names the server software and version. The server checks its credential against the Claude API at startup and every 60 seconds after. And model calls to the Claude API, Foundry and Anthropic-format gateways carry the label cma-server in metadata.user_id. That last one can break things: if your gateway rejects the field, every model call fails.
One deployment, one trust level
Environments don't separate anything. The server API key reaches all of them, and any caller with a role can run sessions in any environment. Anthropic's answer for separate tenants, or for trusted and untrusted code, is a deployment each, with its own database, key and workers.
The MCP allowlist is strict. Hosts must match exactly, wildcards don't work, and the list starts empty. Git is looser. The server limits which branches an agent may push to but doesn't refuse force pushes, so branch protection has to live on your Git host.
Before you move an application
Search your client code for url sources on images and documents, file resources, mcp_oauth credentials, webhook handlers, and anything that calls /v1/messages through the same base URL. All five behave differently on your server, and two of them, the URL sources and the webhook-started workers, give no error.
Then decide who watches what Anthropic used to watch. On your own server a stalled session can mean no worker is polling, and a clean push log can sit on top of a rejected push. You get the same API, and the monitoring for these cases is now yours to write.


