Anthropic Will Let You Run Claude's Agent Stack in Your Own Data Center. Claude Itself Still Runs Outside It, and Scheduling, Webhooks, and Dreams Stay Behind.
Anthropic opened a private beta of self-hosted Claude Managed Agents on November 12. Customers run the Managed Agents server, its PostgreSQL database, and the sandbox workers where agents execute shell and file tools in their own cloud or on premises, with credentials in their own secrets manager and session history in their own database. Model calls still go to Anthropic, Amazon Bedrock, Google's Agent Platform, Microsoft Foundry, or the customer's own gateway. Anthropic's own differences page lists what the self-hosted server does not have: cloud environments, the Messages and Models APIs, the Files API, OAuth MCP credentials, the prebuilt Office skills, scheduled deployments, dreams, and webhooks. Nothing migrates from the hosted service, access is through the account team, and there is no price or GA date. We read the article and the docs.
By T.W. Ghost
Anthropic opened a private beta of self-hosted Claude Managed Agents today, November 12, 2026. The shape: the customer runs the Managed Agents server, its PostgreSQL database, and the sandbox workers where agents run shell and file tools, in their own cloud or on premises. Credentials live in the customer's own secrets manager, session history in the customer's own database. Model requests go out to Anthropic, Amazon Bedrock, Google's Agent Platform, Microsoft Foundry, or a gateway the customer runs. The pitch is control for regulated industries. If a zero-data-retention agreement is in place, Anthropic says it applies to those model requests the way it applies to any Messages API call.
What it is not is Claude on premises. The model still runs outside the customer's network, and Anthropic's own page titled "Differences from the hosted service" lists what the self-hosted server does not do: no cloud environments, no Messages or Models API on the server, no Files API, no OAuth credentials for MCP servers, none of the prebuilt docx, pdf, pptx, and xlsx skills, no scheduled deployments, no dreams, no webhooks. Nothing copies over from the hosted service, including session history. Access is through the account team, and there is no price and no GA date. We read the announcement, the differences page, and the quickstart. Here is what held up.
What Runs Where
| Self-hosted Claude Managed Agents (private beta, November 12, 2026) | |
|---|---|
| You run | Managed Agents server; PostgreSQL database; sandbox workers; credentials in your secrets manager; session history in your database |
| Model requests go to | Anthropic, Amazon Bedrock, Google Agent Platform, Microsoft Foundry, or your own gateway |
| ZDR | If in place, applies to model requests as to any Messages API call (vendor) |
| Server-only features | Choice of inference provider; OIDC single sign-on; your own secret manager; signed agent identity tokens for MCP |
| Not available | Cloud environments (returns 400); Messages and Models API on the server (404); Files API and file resources; mcp_oauth credentials; prebuilt docx, pdf, pptx, xlsx skills; scheduled deployments and dreams; webhooks |
| Migration | None; agents, skills, vault credentials, memory stores, and session history must be recreated |
| Dev stack | Docker Compose, plain HTTP on port 8400 on loopback; docs warn that some Docker Engine releases before 28.3.3 expose it to the LAN |
| Production | Two Helm charts |
| Database; remote access | Postgres 16 or later; Anthropic has no remote access to the server or its logs (press) |
| Access; price; GA | Account team only; not published; not published |
It is orchestration on premises, not a model on premises. Every inference call leaves the network unless the gateway routes to a cloud provider's endpoint, which is still outside the building. For the buyer this pitch targets, that is the first question, and the article answers it plainly.
The feature list runs one way. Scheduling and webhooks are how agents get triggered without a person at the keyboard; dreams are how they consolidate memory. All three stay hosted-only. MCP credentials fall back to static bearer tokens the customer rotates.
Start from zero. No export, no import. An organization with months of hosted sessions, skills, and vault entries rebuilds them.
The docs flag their own dev stack. The Compose quickstart listens on loopback over plain HTTP, and the docs note that some Docker Engine versions before 28.3.3 can expose that port to the local network. Read that line before running it on a laptop in an office.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Private beta announced Nov 12, 2026; access via account team | Verified (claude.com article; docs) |
| Customer runs server, database, workers, credentials, session history | Verified (article) |
| Model request destinations | Verified (article; docs) |
| ZDR applies to model requests | Verified (vendor) |
| Server-only features | Verified (differences page) |
| Not-available list | Verified (differences page) |
| Nothing migrates from hosted | Verified (differences page) |
| Compose on port 8400; Docker Engine warning; two Helm charts | Verified (quickstart) |
| Postgres 16+; no Anthropic remote access | Press (AI Frontier Post) |
| Pricing; GA date | Not published |
Who Should Use It
Ask your account team for it if your rule is that session transcripts, credentials, and sandboxes never leave your estate, and you can accept that the model does.
Skip it for now if your agents depend on schedules, webhooks, dreams, or the Files API, or if you would be moving a hosted deployment; there is nothing to move it with.
The frontier is elsewhere: the hosted service keeps every feature, and the question of where Claude runs when someone else hosts it is the subject of our note on Claude's two Google Cloud versions. Our model comparison has the current standings side by side.
Sources
- Anthropic: self-hosted Claude Managed Agents
- Claude platform docs: self-hosted, differences from the hosted service
- Claude platform docs: self-hosted quickstart
- Pasquale Pillitteri: self-hosted Claude Managed Agents private beta (press)
- AI Frontier Post: Anthropic self-hosted Claude Managed Agents private beta (press)
*Not sure whether your compliance rule is about the model or about the data around it? Take the free 2-minute quiz and get matched. Then read about the feature that stays hosted-only: agent memory and dreams.*