Self-hosted Claude Managed Agents has no public price

Self-hosted Claude Managed Agents has no public price. Anthropic opened the private beta on November 12, 2026. You run the harness, the sandboxes, and the session database. The model call still leaves. Hosted Claude Managed Agents, the service Anthropic runs, stays at $0.08 per session-hour while a session's status is running. A session that stays in that state for all 720 hours of a 30-day month is $57.60 of runtime before any token.

The November 12 article calls the product a new way to run Managed Agents. The release notes the same day say you run the whole service, store agents, sessions, and session history on your infrastructure, and the server sends model requests to the provider you choose, Anthropic by default. The monthly arithmetic, and the platforms that publish a different clock, are 720 hours of Claude agent runtime is $57.60.

What shipped on November 12?

A private beta of the control plane. The hosted harness, which Anthropic has offered since the April 8, 2026 release notes, stays where it is. Self-hosted sandboxes, which the May 19, 2026 notes added, stay a smaller move: you run the tools, Anthropic still runs the agent.

The November 12 article is specific about the new boundary. Every part of the service runs in your cloud or on-premise: the harness, the sandboxes, the credentials, and the session history. You operate the Managed Agents server, its database, and the workers where agents run shell and file tools. Developers keep the same core APIs as the hosted service. Access is a conversation with your Anthropic account team. The docs repeat that sentence on the quickstart, the Kubernetes guide, and the differences page, and they add that the release archive itself is behind the same request.

That is the whole public offer. There is no console button, no list price, and no general-availability date on the pages read November 13.

Where does each piece actually run?

The harness can sit in your network. The weights do not. A model call carries the conversation so far, including prompts, tool inputs, and tool results.

Source: aiagentsinfo.com. The November 12 split is from Anthropic's article and the release notes. The sandbox column is from Choosing an environment and the May 19 release note. Retention and the header rule are from Data and access, read November 13, 2026. The Google name differs by page: the article says "Google's Agent Platform," and the data-and-access page says "Claude Models on Google Cloud (Google managed)."

Anthropic's security claim is narrower than "the prompt never leaves," and it is still a real claim. The data-and-access page says Anthropic has no remote access to the server or its log, Anthropic issues none of the credentials the server accepts, and the server never connects to a worker. Workers have whatever network you give the sandbox. The egress gateway, if you turn it on, connects to hosts you list and can attach stored credentials. Web search and web fetch, once you enable them, run at the inference provider. The server makes no connection of its own for those tools.

Two identifiers do leave when the provider is the Claude API. Each model call can carry metadata.user_id set to the label cma-server, and a User-Agent that names the server software and version. The data-and-access page says neither field identifies your deployment by itself. The API key identifies the organization. Calls to Amazon Bedrock and to Claude Models on Google Cloud (Google managed) do not send metadata.user_id. A gateway of yours that rejects that field will fail every model call that carries it.

Who is the private beta for?

Teams that already have an Anthropic account team, and a security review that fails while Anthropic stores the transcript or can reach the box.

The November 12 article names the work: agents that fix broken builds, test releases, triage alerts, and clear a bug backlog, unattended, with access to internal systems. For that access, some teams want the whole service in their environment. The hosted product is the wrong shape for them, because the overview is explicit: Managed Agents is stateful, transcripts persist until you delete them, and the product is not eligible for zero data retention or a HIPAA BAA. The exception it names is this beta, where you store that data yourself.

It is a poor fit for four other buyers.

• A team that only needs bash and files on its own machines. Self-hosted sandboxes already do that. Anthropic still runs the loop, and tool results still go back so the model can continue. The environment guide is the decision page.
• A team that wants a price. The pricing page sends self-hosted pricing to the account team and prints a rate only for the hosted runtime.
• A team that wants the scheduled deployment and the webhook. The differences page says the server has no routes for scheduled deployments, dreams, or webhooks. You start sessions from your own scheduler and keep workers polling.
• A team that wants to point an existing hosted workspace at the new server. Nothing is copied.

The beta is also a single trust boundary. The data-and-access page says each deployment serves one tenant at one trust level. Environments inside one deployment do not isolate trust. The server API key reaches every environment. Separate tenants, and separate trust levels, each get their own deployment, database, key, and workers.

What do you give up when you leave the hosted service?

The API is the same, and then the differences page spends its length on what is not.

Cloud environments do not exist. A request for one returns HTTP 400. The Messages API and the Models API are not on this server. /v1/messages and /v1/models return 404. File resources and the Files API are rejected. memory_store resources work. Prebuilt docx, pdf, pptx, and xlsx skills are rejected. You upload skills yourself, and a zip upload fails with SKILL.md not found in uploaded files because the server does not unpack archives.

Multiagent orchestration shrinks. The server accepts a multiagent block only when its type is coordinator, with 1 to 20 agents you list. Dynamic workflows, the advisor, and subagents the agent defines itself are absent. The examples that set the type to multiagent_20261001 return 400 on this server.

Web tools are off until the settings file enables them. MCP servers and Git hosts are refused until you list the exact hostnames. The MCP result cap defaults to 100 KiB, and the rest of a longer result is dropped. On the hosted service, a long result can be saved to a file in a cloud sandbox. An image or document sent as a URL is accepted and the model never sees it. Send base64.

The Claude Console at platform.claude.com does not show your server. You use the console the server itself serves, on the same address as the API, with no credential required to load the page. Doing anything in it requires your identity provider. That is worth a firewall rule before the Helm install, not after.

How does this compare with AgentCore and with OpenAI's hosted sandbox?

Those products host a place for tools to run. They do not ship Anthropic's harness, and they do not take Anthropic's $0.08.

Amazon Bedrock AgentCore is a microVM you put your own agent in. The November 12 runtime change, the V1 and V2 rates, and the 8-hour microVM lifetime are AgentCore runtime V2 is now the default. A fully busy hour at 1 vCPU and 2 GB on the V2 consumption rates is $0.1614 before system overhead. The same busy hour of hosted Claude runtime is $0.08, and the model tokens are extra on both invoices. AgentCore can scale CPU to zero during I/O. Claude's $0.08 accrues for the whole time status stays running, including the model call, and drops to zero when status is idle. The clocks are different products.

OpenAI's Agents API, read November 13, bills model tokens at API rates and OpenAI-hosted sandboxes at the standard container rates. The sandbox page's default size is medium, 2 vCPU and 4 GB. The pricing page prices a 4 GB container at $0.12 per 20-minute session. That is a sandbox fee with a model fee beside it. It is not a Managed Agents session. Codex Cloud is a third thing: the Codex help article, retrieved the same day, says standard Codex Cloud environments have no separate VM charge at launch, and usage counts against the plan. A Codex seat is not this beta, and it is not the $0.08.

If the question is "where can the loop run," the November 12 answer is: on a server you operate, talking to a model provider you name. If the question is "where can the tools run," you have had an answer since May 19, and AgentCore and the OpenAI sandbox are other answers with their own meters. The always-on cost note prices those meters for 720 hours. Seat prices for Claude Code and Codex, which are not this runtime, stay on Claude Code vs Codex pricing.

What is the hosted bill, and which limits are printed?

$0.08 per hour of running. Tokens are the model card. A one-hour Opus 5.5 example on the pricing page is $0.58 with 50,000 input tokens and 15,000 output tokens, or $0.428 when 40,000 of those input tokens are cache reads.

Source: aiagentsinfo.com. Rates and the two worked examples are from the pricing page, read November 13, 2026. Each 720-hour cell is the published one-hour total times 720. November 2026 has 30 days, so 720 is 30 × 24. Idle, rescheduling, and terminated time are outside the runtime meter. Web search inside a session is an extra $10 per 1,000 searches. US-only inference is 1.1× on tokens. The Batch API discount does not apply.

The pricing page does not print a separate container fee on top of session runtime. It also does not print a waiver of the $0.08 when you bring your own sandbox. The self-hosted service, the one where you run the server, is the row with no number.

Limits that are printed, and that change a design review:

The quickstart is blunt about the Compose stack. It serves plain HTTP, and agent tools run in containers on that host, so the machine should hold nothing you need to protect. Production is the Helm path. The upgrade guide is the other operational fact: a minor upgrade can change the database schema, the dump holds every conversation in plain text, and the guide tells you to encrypt the backup. Plan for that before the first session worth keeping.

What should you do this week?

Ask your account team if the review in front of you is about the transcript and the server, and you can staff a Kubernetes cluster and a PostgreSQL 16 database. Read the differences page before anyone writes "lift and shift" in a ticket. Budget the hosted $0.08 for whatever you run while you wait, and treat idle time as a status you can see, because that status is the difference between $57.60 and $0 of runtime.

Wait if the requirement is only "tools on our network." Self-hosted sandboxes are the feature that matches that sentence, and they have been in the release notes since May 19. Waiting is also the right call if you need webhooks, scheduled deployments, or a copy of the agents you already built. Those are absent, and the docs say so in a table rather than a footnote.

The numbers in the table are licensed CC BY 4.0. Credit AI Agents Info if you reuse them. They are arithmetic on Anthropic's published hour, not an invoice.

Frequently asked questions

Is self-hosted Claude Managed Agents generally available?

The November 12, 2026 article and the platform release notes call it a private beta. You request access, and the release download, from your Anthropic account team. Hosted Managed Agents is a separate beta. The overview says that one is enabled by default for all API accounts, and every endpoint still sends the managed-agents-2026-04-01 beta header.

Does self-hosting keep the prompt inside your network?

The transcript stays in your database. The model call sends the conversation so far, including prompts, tool inputs, and tool results, to the provider you configured. The default is the Claude API. The data-and-access page says the server connects to Anthropic only in that case, and that Anthropic has no remote access to the server or its log.

How much does it cost?

Self-hosted pricing is not on the pricing page. The page says to contact your account team. Hosted runtime is $0.08 per hour while status is running. Seven hundred twenty of those hours cost $57.60 before tokens. Anthropic's own one-hour Opus 5.5 example, scaled by the same 720, is $417.60 uncached and $308.16 with the cache split the page prints.

Can you migrate a hosted workspace?

The differences page says nothing is copied, session history included. You create the agents, skills, credentials, and memory stores again. Scheduled deployments and webhooks are not on the server.

Who should apply this week?

A team with an account team, a cluster they will actually operate, and a review that fails while Anthropic holds the transcript. A team that only needs the sandbox on its own machines should use self-hosted sandboxes and skip the beta.