Self-hosted Claude Managed Agents: architecture, security and deployment
What self-hosted Claude Managed Agents actually keeps inside your network, how credentials and sandboxes work, and what to test before moving an enterprise agent.
BY LUKAS HOFFMANN · FDEINTERVIEWS EDITORIAL · UPDATED NOVEMBER 15, 2026 · 5 MIN READ
PRACTICE THIS:Practise enterprise agent security ·Learn harness engineering ·Design production AI agents
Self-hosted Claude Managed Agents puts the agent service, session database and execution sandboxes on infrastructure you operate. Claude inference still happens at the configured model provider. Choose it when you need control over stored sessions or provider routing and can operate the service yourself. If your only requirement is running tools beside private systems, self-hosted sandboxes may be sufficient. Anthropic's deployment overview distinguishes those options explicitly.
For an FDE, the first design question is where each copy of customer data goes. A box labeled “our cloud” cannot answer that by itself.
What do you actually host?
The service runs the API and agent loop. Workers execute shell and file operations in your sandboxes; PostgreSQL stores the service's persistent state. Your applications authenticate to the service, which separately authenticates to the inference provider. Anthropic documents its server under a specific software license, with download access through the account team. Treat access and support terms as deployment prerequisites.
| Deployment | Agent service | Tool execution | Operational consequence |
|---|---|---|---|
| Hosted service and cloud sandboxes | Anthropic | Anthropic | Least infrastructure to operate |
| Hosted service, self-hosted sandboxes | Anthropic | Your infrastructure | Private tool execution; hosted orchestration |
| Self-hosted Managed Agents | Your infrastructure | Your infrastructure | You own service operation and session storage |
Our BYOC deployment concept explains why these ownership boundaries matter. In the figure, follow the amber model connection out of the customer boundary. The local database does not remove that connection.
A private sandbox still needs enforced boundaries
Imagine a support agent that reads incident reports and proposes configuration repairs. Its report can contain attacker-controlled instructions. A shell with production credentials turns a reading task into a much larger capability.
Anthropic's security guidance separates model instructions from controls enforced by the surrounding system. For this example, start with read-only incident access and a disposable workspace. Put the repair action behind a tool that checks the requesting identity, target environment and permitted change. Require separate approval for a production write. This follows the same reasoning as defense in depth for an agent with write authority.
The worker is not itself a sandbox. Its shell inherits the permissions and network reach of the container or VM that runs it. Test isolation with a fixture containing a fake secret, a denied destination and a forbidden file path. The expected result is a concrete denial, not an agent politely declining to use them.
With an egress gateway configured, a placeholder can be exchanged for a stored credential outside the worker. But the destination still receives an authenticated request. A read-only token and a token that can delete records have very different consequences. Bind credentials to the narrow destination and operation, and verify that the worker cannot bypass the gateway. The configuration reference is the place to check the exact settings for your release.
Map data leaving the deployment
The server sends the conversation, including tool inputs and results, to its inference provider. MCP and Git connections, identity services, secret managers and configured telemetry destinations create additional paths. Workers have their own network access. Anthropic's data and access documentation describes these separately.
Build a small inventory before the security review:
| Copy of data | Decide before release |
|---|---|
| Model request | Provider, region where applicable, permitted fields and contractual retention |
| Session database | Tenant separation, backups, deletion and restore permissions |
| Sandbox workspace | Lifetime, mounted files and treatment of failed sessions |
| Telemetry export | Redaction, sampling, operator access and retention |
Traces and metrics are opt-in through OpenTelemetry configuration. Session history is a separate store; a trace retention policy does not delete it. Deleting a session also does not erase database backups or material copied elsewhere. Exercise the full deletion path with a synthetic customer record, then search each store for its identifier.
This is the applied version of agent observability: retain enough evidence to diagnose an action without collecting every sensitive field by default.
Plan for feature differences before migration
The self-hosted API has documented exceptions. As checked on November 15, 2026, it does not provide the hosted Files API, scheduled deployments or webhooks. Custom skills need uploading; the prebuilt office-document skills are not simply present. Hosted session history is not copied into the new deployment. The differences reference also limits multiagent orchestration to predefined coordinator agents.
Convert that list into acceptance cases. An application that starts workers only from a webhook needs a different wake-up path. A skill that expects a file resource needs an explicit workspace transfer. An agent that discovers new subagents at runtime needs a redesigned delegation contract.
Keep a versioned manifest of agent definitions, skills, tool permissions and required features. Recreate these in a fresh deployment and compare the result with the manifest. A successful health endpoint says little about whether a critical skill or credential was migrated.
Run a deployment rehearsal with one useful task
Use Anthropic's Docker Compose quickstart in an isolated environment with synthetic data. It starts the server, database and sandbox workers. Do not reuse the service API key as the provider credential: they authenticate different connections. This article reviews the documented operator workflow; it does not claim a live private-beta deployment test.
For the incident agent, rehearse five conditions:
- A normal incident produces a proposed repair with the evidence it used.
- A forbidden network destination fails even when the incident text requests it.
- A stopped worker leaves a diagnosable pending tool call and can resume safely.
- A restart retains sessions through the configured database; a backup restores them.
- Repeating an approved action cannot apply the repair twice.
Record the expected event and operator response for each failure. The sandboxing concept and credential-isolation question provide the deeper design work. Release the first read-only workflow only after these checks have an owner and a repeatable result.
Turn it into offers. Work the real questions and concepts this maps to:
FAQ
Anthropic documents a self-hosted Managed Agents server and sandboxes running on your infrastructure. Access to the server download requires contacting its account team. Check the current license and feature differences before planning a migration.
No. The service and sandboxes run on your infrastructure, while model calls go to your configured inference provider. Those requests include conversation content and tool results, so local session storage does not imply local inference.
The documented egress gateway can substitute stored credentials outside the sandbox. That depends on configuring the gateway and preventing bypass; credentials you put directly into a worker's environment or files remain accessible to that worker.
No. The server emits traces and metrics only after you configure the OpenTelemetry environment variables. You own collector access, retention and the treatment of sensitive trace data.
Discussion (3)
The useful first artifact is a data-flow inventory. Session history, model input, a repository checkout and an exported trace can hold the same customer material under four different retention policies.
A credential proxy removes the need to reveal a secret to the worker. The service receiving the authenticated request still needs to check what that identity may do. A valid credential alone is too broad a permission test.
A migration rehearsal should include a session waiting on a tool call. Model turns can work while no worker is polling, which makes a broken worker deployment look like a model that stopped making progress.
