Self-hosted Claude Managed Agents changes who operates the agent service. It does not mean Claude's model weights run on your server, that every hosted feature is available, or that a workflow becomes compliant merely because you control its database.
Anthropic listed the release on November 12, 2026. This guide explains the deployment decision and proposes a small acceptance-test plan. It is documentation-based analysis, not a Nerova installation report, security certification or performance benchmark. Official release notes.
What self-hosted means here
There are three different arrangements: Anthropic hosts the service and sandboxes; you host only the sandboxes; or you host both the service and sandboxes. In the third arrangement, inference still goes to Anthropic or another configured provider. The service includes the API, agent loop and session storage. Access to the server download is requested through the Anthropic account team, and the software has its own license.
You operate and update the server, database and execution environments. Durable storage uses PostgreSQL 16 or later; the default in-memory setup loses its state on restart. The worker's shell inherits the permissions and network access of its execution environment. Treat isolation as an operational responsibility, not a property supplied by the word “worker.” Self-hosted overview.
Start with the requirement, not the hosting preference
Write down the business reason for changing deployment. “Our approved workflow records must be stored in a database we administer” is testable. “We want everything private” is too vague to evaluate. Specify the data, allowed destinations, retention obligations, users and actions involved.
Then ask whether moving only tool execution would satisfy the requirement. Full service ownership adds work. Assign a named owner for upgrades, incident response, recovery and cost review before choosing it. A team that cannot staff those responsibilities has not solved an operations problem by changing the address of an API.
Data ownership is not the same as no data leaving
Model calls include conversation content, including tool inputs and results. Provider retention terms apply to those calls separately from your database and telemetry. Server allowlists do not constrain the network access of worker commands. Distinct trust levels require separate deployments, not just different environments.
Deleting a session does not erase its copies in backups, exported telemetry, memory stores or retained workspace volumes. Your data map must cover those locations as well as inference. Official data and access documentation.
For a synthetic pilot, use fictional records and deliberately recognizable test markers. Record where each marker is expected to appear and where it must not appear. Inspect only authorized logs and destinations. Absence from one application log does not establish absence from backups or downstream services.
API compatibility is not workflow parity
The self-hosted differences page documents no automatic transfer of hosted agents or session history. Hosted cloud environments, the Files API, scheduled deployments, dreams and webhooks are unavailable. A tool call can wait without error when no worker is polling. URL-based image or document inputs can be accepted without being shown to the model. These are important reasons to test observed behavior rather than stopping at a successful HTTP response. Hosted versus self-hosted differences.
Do not migrate an entire portfolio at once. Select one bounded workflow whose dependencies you understand. List its triggers, inputs, tool actions, expected outputs, state transitions and recovery path. Mark every required capability as proved, unsupported or not yet tested. A missing capability should become an explicit design decision, not an invisible fallback.
A seven-check migration gate
The following is an original evaluation plan, not a claim that these tests have been run. Use disposable infrastructure, synthetic data and mock external writes. Establish the permitted actions before running any case.
| Requirement | Safe test | Evidence to collect | Decision |
|---|---|---|---|
| Durable work | Record a synthetic checkpoint, restart the test service and retrieve it. | Exact record and state before and after restart. | Do not promote if required state disappears. |
| Worker availability | Pause the isolated worker during a harmless tool request. | Visible pending state, actionable alert and recovery after restoration. | A silent stall is not an acceptable completion signal. |
| Attachment understanding | Provide a synthetic document containing a unique fact through the intended input path. | A correct answer grounded in that fact, not merely an accepted upload. | Block the workflow if required input is unseen. |
| Access separation | Use two separate test deployments for distinct trust levels, with test identities and fixtures carrying explicit permissions. | Allowed access succeeds and forbidden cross-boundary access fails. | Investigate unexpected access without broadening permissions. |
| Network limits | Exercise approved and deliberately disallowed synthetic endpoints. | Expected results and authorized egress evidence for each execution component. | Do not assume a server setting covers workers. |
| Deletion policy | Delete an eligible test session and inspect the locations named in the retention plan. | What was removed, what remains and its authorized retention owner. | Resolve unexplained copies before using sensitive records. |
| Safe retries | Simulate a timeout after a mock destination records a write. | Reconciliation finds the existing operation instead of creating a second one. | A resumed session must not imply permission to repeat an uncertain write. |
Keep an evidence row for the software version, model/provider configuration, test date, owner and result. A screenshot of a successful answer is not a substitute for checking the intended destination and state. A failed test should record the observed result, not be relabeled as passed because an operator manually repaired it.
Compare the full operating cost
Ask for applicable commercial terms rather than inventing a public server-license price. Build a budget that separates inference, compute, storage, database operations, observability, upgrades and human review. Include expected failure recovery and maintenance time. A lower model bill can coexist with a higher overall operating bill.
For illustration only, suppose one controlled evaluation costs $120 and yields 80 accepted tasks. That is $1.50 per accepted task. A second configuration costing $100 but yielding only 50 accepted tasks costs $2 per accepted task. These are hypothetical figures, not vendor prices or Nerova results. Define acceptance before the run and report excluded or unfinished cases rather than quietly removing them.
Compare equivalent tasks, permissions and input quality. Record latency and failure classes alongside cost. A different provider, prompt, runtime and permission policy changed at once makes it difficult to identify why an outcome changed.
Make rollback a product requirement
Before a pilot, decide which work can safely stop, which operations need reconciliation and which records must remain available. Keep a versioned configuration and a way to route new test work to the prior approved path. Do not assume that copying a prompt reconstructs a live session or its external effects.
For an uncertain write, inspect the authorized destination or its operation receipt before retrying. For a missing capability, use an explicitly designed alternative or exclude that workflow from the pilot. Do not disable access controls, expand network access broadly or switch to real customer data merely to get a green demonstration.
When this option is worth evaluating
Self-hosting is worth investigating when a specific ownership or infrastructure requirement justifies the operational work and your required workflow survives the acceptance checks. It is not automatically cheaper, safer or easier than hosted service use.
If you are still deciding what an agent should do, begin with the task and its authority. Tool-calling boundaries and a goal-evaluation worksheet can help define that contract before infrastructure choices dominate the project. For the original hosted-service context, see our Claude Managed Agents overview.