On this page · 12 sections
- What launched on 12 November 2026
- What you run, piece by piece
- Which model provider your calls go to
- What still leaves your network
- What you lose against the hosted service
- Your security and operations load
- The licence terms to read before you sign
- What it costs
- India-specific considerations
- A decision checklist
- How eCorpIT can help
- References
Summary. On 12 November 2026 Anthropic launched self-hosted Claude Managed Agents: you run the whole Managed Agents service, not just the sandboxes, in your own cloud account or data centre, and agents, sessions and session history are stored there. A deployment is the cma-server process (API plus agent loop), a PostgreSQL 16 or later database, workers that run the agent's shell and file tools in your sandboxes, and a console where people sign in through your identity provider. Model calls go to the Claude API by default, or to Amazon Bedrock, Claude Models on Google Cloud (Google managed), Microsoft Foundry or your own gateway, and every call still carries the conversation so far, including tool inputs and results. Access runs through the Anthropic account team, and Anthropic publishes no self-hosted price; the hosted service costs $0.08 per running session-hour plus tokens. In return you take on server security, sandbox isolation, every upgrade and licence deadlines as short as eight hours, and eight hosted features are missing. Our verdict: worth it when transcripts must sit in your own database or model calls must go through your own cloud contract; otherwise self-hosted sandboxes or the hosted service are less work.
About this article. This is for platform, security and architecture leads who run, or plan to run, agents on Claude Managed Agents and have a data-location, cloud-commitment or audit reason to bring the service in-house. On 20 November 2026 we checked every component, setting, port, date and price against Anthropic's self-hosted documentation (overview, differences, inference provider, data and access, quickstart, Kubernetes, upgrade, single sign-on and console pages), the Claude Platform release notes, the pricing and data retention pages, and the server licence agreement. Rupee figures use European Central Bank reference rates. Anthropic can change features and terms between releases, this is not legal advice, and no vendor paid for inclusion. Researched with AI assistance and checked against the primary sources listed in the references.
What launched on 12 November 2026
Managed Agents has two parts you can host: the sandboxes where an agent's tools run commands and edit files, and the service itself, meaning the API, the agent loop and the storage for agents, sessions and history. Self-hosted sandboxes arrived on 19 May 2026 and moved only the first part to you; the 12 November release moves the second part as well. Anthropic's overview names the three options by what you host.
| Option | Sandboxes | Service and storage | Model |
|---|---|---|---|
| Cloud sandboxes (default) | Anthropic | Anthropic | Anthropic |
| Self-hosted sandboxes (since May 2026) | You | Anthropic | Anthropic |
| Self-hosted Managed Agents (since 12 Nov 2026) | You | You | Anthropic or a provider you choose |
Your server exposes the same API as the hosted service, so the SDKs and the ant CLI work against it once you point them at your server. Nothing moves across automatically, though. Agents, skills, vault credentials, memory stores and session history all have to be created again on your server. If you run agents on the hosted service today, plan a rebuild, not a migration.
Anthropic gives two reasons to choose this option: you must store agents, sessions and session history yourself, or model calls must go to Bedrock, Google Cloud, Foundry or your own gateway. If you only need tools to run inside your network, Anthropic points you to self-hosted sandboxes, where it keeps running the service. Our comparison of agent session isolation across AWS, Azure, Google and Anthropic covers that lighter option.
What you run, piece by piece
The overview lists six parts of a deployment.
- Server. The
cma-serverprocess serves the API, runs the agent loop and stores agents, sessions and history. It calls the inference provider when the agent needs the model and never runs the agent's shell or file tools itself.
- Database. PostgreSQL 16 or later. Without one the server keeps everything in memory, and a restart wipes it.
- Workers. Programs that poll the server for work and run the shell and file tools in your sandboxes, started as an
ant beta:workercommand or from an SDK.
- Your applications. Your code, calling the server through the SDK or the CLI.
- Console on your server. A web page at the server's own address where people sign in through your identity provider. It is separate from the Claude Console, which cannot see anything on your server, and it does not accept the server API key.
- Admin CLI. The
cma-server-admintool creates environments and their keys, and runs on Linux (x86-64 and Arm) and on macOS with Apple silicon.
Three credentials hold it together: a server API key that your applications and CLIs present, an inference credential that only the server holds for model calls, and an environment key per environment that workers use to pull work. Applications and workers never need the inference credential.
Two install paths
The quickstart runs everything on one machine with Docker Compose: the server, PostgreSQL and workers that start each session in its own container. It needs at least two CPUs, the latest Docker Engine with the Compose plugin, openssl and the ant CLI, and Anthropic labels it development only because it serves plain HTTP.
For production, Anthropic recommends the two Helm charts in the release archive. The Kubernetes guide asks for Kubernetes 1.30 or later with the Agent Sandbox controller, a sandboxed container runtime (gVisor nodes on Google Kubernetes Engine and Amazon EKS, Pod Sandboxing on Azure Kubernetes Service), a network plugin that enforces NetworkPolicy, a PostgreSQL 16 database outside the cluster, and a registry you push to. You build three images yourself (server, dispatcher and sandbox), and the server chart runs two servers behind a Service.
Which model provider your calls go to
The inference.provider setting decides where model calls go.
| Setting | Model calls go to | Server authenticates with |
|---|---|---|
| anthropic (default) | Claude API | Claude API key |
| bedrock | Amazon Bedrock | AWS default credential chain |
| vertex | Claude Models on Google Cloud (Google managed) | Google Application Default Credentials |
| foundry | Microsoft Foundry | Foundry API key or Microsoft Entra ID |
| gateway | A model gateway you run | A key or token your gateway accepts |
Agents keep naming models by their Claude API IDs, such as claude-opus-5-5, and the server translates the ID for the provider, so one agent definition works under any provider that serves the model. On Bedrock the server adds the us.anthropic. prefix by default, which names a US cross-region inference profile; you change the prefix or map IDs one by one if your account uses other profiles. The gateway option accepts three request formats (Claude Messages, Bedrock InvokeModel and Google's rawPredict) and supports short-lived tokens through an OAuth exchange that your own callback signs, or Entra ID tokens issued to the server's Azure identity.
Under every provider except the Claude API, the server needs no Claude API key. That matters for teams whose AI spend is committed to one cloud: the agent loop, history and tools stay in your account and the model is bought through the same cloud contract.
Features that depend on the provider
The provider choice changes what the agent can do, and some of these gaps return no error.
- Web search works only under the Claude API, Google Cloud and Foundry; web fetch only under the Claude API and Foundry. Elsewhere the agent saves and runs, but the model is never offered the tool.
- The
inference_geosetting works only under the Claude API; the server refuses to save an agent that sets it under any other provider.
- Fast mode depends on the model catalogue: by default four Opus models qualify, Foundry lists only Claude Opus 5.5, and Google Cloud lists none.
- Under the Claude API, the server asks for server-side fallback by default, so a request declined by a safety classifier can be answered by another, possibly less capable, Claude model that Anthropic picks. You turn this off with
inference.model_fallback: false. Under other providers such a request ends the turn with a refusal.
Both web tools are also off on every self-hosted server until you switch them on in the settings file. If you have tuned web domain filters on the hosted service, our note on Managed Agents web tool domain allowlists explains the agent-level lists, which still apply once the server-level switch is on.
What still leaves your network
Self-hosting moves storage, not inference. For every model call, the server sends the conversation so far, including prompts, tool inputs and tool results, to the inference provider. If that provider is the Claude API, each call also carries up to three headers with the session, agent and environment IDs. The IDs contain no names or content but do encode when your server created them, and no setting turns the headers off. Under any other provider they are never sent.
There is a smaller second channel. Requests to the Claude API carry a User-Agent naming the server software and version, and calls to the Claude API, Foundry and Claude-format gateways carry the fixed label cma-server in metadata.user_id, which Anthropic can use to count self-hosted usage; Bedrock and Google Cloud calls never carry it. The licence adds that you must not remove, block or falsify these usage signals. With the Claude API as provider, the server also checks its credential at startup and every 60 seconds, without running inference.
What does not leave: Anthropic states it has no remote access to the server or its log, issues none of the credentials the server accepts, and that the server connects to Anthropic only when the Claude API is the provider. The documented outbound list is short: the provider, the database, credential issuers, the MCP servers and Git hosts you allow, your OpenTelemetry collector, your identity providers and your secret manager. Anthropic suggests comparing that list with your egress proxy or firewall log, which is what an auditor will ask for.
Retention and zero data retention
Anthropic's data retention page lists hosted Managed Agents as not eligible for zero data retention, because Anthropic stores session transcripts until you delete them; self-hosted Managed Agents is the stated exception. On your server, sessions stay in your database until you delete them, and the server deletes nothing by age. What the model provider keeps falls under your agreement with that provider. One trap from the upgrade guide: a database dump holds every conversation in plain text, which is why Anthropic's commands encrypt backups with the age tool.
What you lose against the hosted service
Anthropic's differences page is the document to read before you commit. The main gaps:
| Hosted feature | On your server | Workaround Anthropic gives |
|---|---|---|
| Cloud environments | Only self-hosted environments; cloud requests return 400 | Run your own workers for every environment |
| Messages API and Models API | Not served; calls return 404 | Send them to the Claude API or your provider directly |
| Files API and file session resources | Rejected | Put files in the worker's directory; send images and PDFs as base64 |
| OAuth MCP vault credentials | Rejected, no refresh flow | Store a static bearer token and rotate it yourself |
| Prebuilt docx, pdf, pptx and xlsx skills | Rejected | Upload them as custom skills from the public skills repository |
| Scheduled deployments | No routes | Start sessions from your own scheduler, such as cron |
| Dreams | No routes | No replacement |
| Webhooks | Never sent | Follow the event stream and keep workers polling |
Several other behaviours change quietly. Multiagent orchestration accepts only a coordinator with 1 to 20 listed subagents, with no advisor or dynamic workflows. MCP servers work only for hosts you list exactly, and the list is empty by default. Git repositories need an allowlist and a list of branches agents may push to, and the server does not block force pushes, so protect branches on the Git host. Images or documents sent by URL are accepted but never reach the model, with no error. MCP results above 100 KiB are cut and the rest is lost. A session's model turns keep running when no worker is polling, and the first tool call then waits silently.
What you gain
The server also has four features the hosted service lacks: the provider choice above, single sign-on through any OpenID Connect provider with roles drawn from group claims, storage of new vault secrets in your own secret manager, and signed short-lived agent identity tokens that your MCP servers or gateways can verify. Single sign-on gives three roles: admin, developer and user. Developers and users see only what they created and cannot share resources, so a team that wants shared agents has to manage them as admin or through the server API key.
Your security and operations load
The overview is blunt about isolation: the agent's bash tool is a real shell with the worker's permissions and network access, and the worker does not isolate it. The Compose stack and the sandbox Helm chart give each session's worker its own container or pod; if you start workers another way, Anthropic says to run them in a disposable container or VM holding nothing you need to protect.
Tenancy is per deployment. Environments inside one deployment do not separate trust levels, the server API key reaches every environment, and any caller with a role can run sessions in every environment. Separate tenants, and separate trust levels such as trusted and untrusted code, each need their own deployment with its own database, key and workers. Defaults are conservative: the API listens on 127.0.0.1 port 8400, and two optional listeners are off until you turn them on. A Prometheus endpoint on port 9464 asks for no credential, and the egress gateway listener on port 8410 binds every interface, speaks plain HTTP and returns stored credential values, so only the gateway should reach it. Anthropic also asks for Docker Engine 28.3.3 or later, because some earlier releases expose loopback-published ports to the local network.
Upgrades are yours. A patch release keeps the database schema; a minor release changes it and needs a backup first; skipping a minor release or a major version bump means stopping the servers inside a maintenance window. On Kubernetes, patch and single-step minor upgrades run without downtime, while a Compose stack is down during every upgrade.
The licence terms to read before you sign
The licence for the server is dated effective 12 November 2026, and for each customer its terms start on the day the binary is first downloaded. Six points stand out for architects and procurement:
- Two permitted uses. You may run the binary as an internal deployment for your own organisation and affiliates, including inside your own products, or as a managed service for your customers, and Anthropic may limit you to one of them.
- Binary only, no changes. The licence is royalty-free but bars modifying the code, reverse engineering it, or bundling it into another product.
- Third-party models. You may not point the software at a non-Anthropic model except where the documentation expressly allows it.
- Updates on a clock. You must deploy a safety update within one business day of delivery, disable a feature or roll back a version within eight hours of a written safety notice, and make reasonable efforts to apply other updates within 30 days. Anthropic has no obligation to support any version.
- Incident notice. You must tell Anthropic within 48 hours of confirming an agent incident, as defined by thresholds the parties agree in writing.
- Model terms unchanged. Model use through your deployment stays under your existing agreement with Anthropic, with its pricing, usage policies and data terms, and you are responsible for all data the deployment processes.
The eight-hour rollback duty is the clause most teams will underestimate. It assumes someone on call who can change a production deployment at short notice, every day of the year.
What it costs
Anthropic's pricing page bills the hosted service on tokens at standard model rates plus $0.08 per session-hour, counted only while a session is running, and directs self-hosted customers to their account team. Model rates on 20 November 2026 include $2 input and $10 output per million tokens for Claude Sonnet 5.5, and $4 and $20 for Claude Opus 5.5.
To see where the runtime fee sits, take an illustrative team of 20 agents running six hours a working day for 22 days: 2,640 session-hours. At the ECB rates for 19 November 2026 (93.30 rupees per US dollar, crossed from the euro rates and rounded to two decimals), the hosted bill looks like this.
| Line | Assumption | Monthly cost |
|---|---|---|
| Session runtime | 2,640 hours at $0.08 | $211.20 (about ₹19,705) |
| Tokens on Sonnet 5.5 | 1M input and 100,000 output per session-hour, no caching | $7,920.00 (about ₹7,38,936) |
| Total | Runtime is 2.6% of the bill | $8,131.20 (about ₹7,58,641) |
The arithmetic: 2,640 x $0.08 = $211.20; each session-hour costs $2.00 + $1.00 = $3.00 in tokens, so 2,640 x $3.00 = $7,920.00. Self-hosting removes, at most, a line worth a few per cent of a typical bill, while adding servers, a PostgreSQL database, a Kubernetes cluster with sandboxed nodes and on-call staff. The business case is control and data location, not savings. For guard rails on the token side, see the gaps in Claude session budgets.
India-specific considerations
For Indian organisations the attraction is clear: agent definitions, transcripts and tool outputs can live in a PostgreSQL database in an Indian cloud region or your own data centre, under your own access logs and backups. That helps with internal data-location policies and with showing a regulator or client exactly where conversation records sit. It does not keep model calls in India by itself. The full conversation still goes to whichever provider you configure, and inference_geo works only when that provider is the Claude API. Check with your cloud provider which regions serve the Claude models you need before you promise Indian processing to anyone.
Running the store yourself also makes deletion, backup and access duties visibly yours, and the licence says so in contract terms: the licensee is responsible for all data a deployment processes. This is general information, not legal advice.
A decision checklist
- Name the reason. Write down whether you need transcripts in your own database, model calls through your cloud, or both. If neither, stay hosted or use self-hosted sandboxes.
- Read the differences page against your agents. Flag any use of webhooks, scheduled deployments, files, prebuilt document skills or OAuth MCP credentials.
- Pick the provider and test the tools. Confirm web search, web fetch and fast mode behave as you expect under that provider.
- Design tenancy first. One deployment per tenant and per trust level, with single sign-on mapped to roles.
- Staff the clock. Agree who can apply a safety update within one business day and a rollback within eight hours.
- Ask for the price in writing. Self-hosted pricing comes only from the account team.
- Pilot on Compose, ship on Kubernetes. Use the quickstart to learn, then the Helm charts with sandboxed nodes for production.
Teams already running self-hosted sandboxes can reuse their workers, because the worker software is the same, but a worker that starts on a webhook will not work against your own server. Our piece on memory stores in self-hosted sandboxes covers the worker behaviour that carries over.
How eCorpIT can help
eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our AI and machine learning services team can assess whether self-hosted Managed Agents fits your data and cloud requirements, then plan the Kubernetes, provider and single sign-on setup. Talk to us at /contact-us/.
Last updated: 20 November 2026.