What does an agent cost when your firm runs the agent service itself? That is the question under Anthropic’s November 12, 2026 launch post for self-hosted Claude Managed Agents, now in private beta. The whole service moves into the customer’s own cloud or building: the harness, the sandboxes, the credentials and the session history. A Canadian firm in a regulated trade gets the control it has been asking vendors for. It also takes on a server, a database and a set of workers to run, and the announcement publishes no price for any of it.
The direct answer: Self-hosted Claude Managed Agents lets a firm operate the full agent service on its own infrastructure, with session history in its own database and credentials in its own secrets manager. Model requests still leave the building and go to a provider the firm chooses, and those requests are billed as model use. The saving on the table is control over records, not a smaller model bill. A firm should budget this as an operations commitment with a model bill attached, and get the missing price in writing before it plans a rollout.
- What actually moves into your building
- What still leaves, and what it costs
- The 3 questions to get answered in writing
What actually moves into your building
Managed Agents is the set of APIs Anthropic sells for putting long-running agents into production. It supplies the harness that calls the model, the sandboxes where agent code runs, credential handling and session history. The hosted version runs all of that on Anthropic’s platform. The self-hosted version hands the whole stack to the customer. Your team operates the server, its database and the worker machines where agents use shell and file tools. Your developers build agents with the same core APIs as the hosted service.
The control gains are concrete. Session history sits in your own database. Credentials can sit in your own secrets manager, and the design keeps them out of the sandbox: a proxy outside the sandbox adds the credential to a request, so the code the agent runs never holds the key. Network access can be limited to a list of allowed hosts, so an agent that fixes builds can reach your git host and your package registry and nothing else. Every session leaves a full record of tool calls and results, exported to the monitoring tools your firm already runs.
For a Canadian firm under a privacy law or a client contract with data terms, those gains answer the questions a regulator or a customer will ask. Canada’s critical cyber law skips hospitals as the UK warns of AI ransomware showed how uneven the legal map still is by sector. A firm that can point to its own database, its own vault and its own session record holds a stronger answer than a firm pointing at a vendor’s promise. One Claude Code setting leaked 1,379 passwords and keys. The check takes 10 minutes showed the cost of getting the credential half wrong. Keeping credentials out of the sandbox is the architectural answer to that failure.
What still leaves, and what it costs
The model does not move. Each request goes where the firm sets it to go: to Anthropic directly, to Amazon Bedrock, to Google’s Agent Platform, to Microsoft Foundry, or through the firm’s own gateway. The full conversation travels with that request, because the model needs it to answer. Self-hosting decides where the records live and who operates the machinery. It does not make the work on-premise in the way a board member may picture on first hearing, and a firm should say that plainly in its own approval paper before a director repeats the shorter version.
Now the money. The announcement publishes no price for the self-hosted service, no general availability date and no public sign-up. Access runs through the firm’s Anthropic account team. Model use is still model use, billed under the firm’s model arrangement with whichever provider it routes to. So the budget shape changes in a specific way: a hosted subscription line shrinks or vanishes, an operations line grows in its place, and the model bill stays. The operations line is your own people’s time: running the server, patching it, watching the workers, holding the database. None of that arrives in a vendor quote, because the vendor is not doing it.
That shape is familiar from any contract where the headline term hides the working term. ChatGPT Wallet lets an agent pay. Approval is the whole contract made the point for agent spending this week: the rule around the tool decides the cost. A judge says ’20x’ has to mean 20x. Read your AI plan like a contract made it for plan limits: the words in the agreement are the product. Here, the words that matter are the ones the announcement does not carry. No price is a fact a buyer can use, as long as the buyer treats it as a gap to close and not a zero to budget.
The 3 questions to get answered in writing
Ask first what the service costs, in full: any fee for the server software, any support terms, and what happens to the price when the beta ends. Ask second who carries a failure at 2 a.m.: your team runs the server and the database, so the incident rota, the patch window and the rollback path are yours to staff, and the account team should confirm in writing what help Anthropic provides when a self-hosted server fails mid-run. Ask third what leaves under your data terms: which requests travel to which provider, and whether your zero data retention terms, if you hold them, cover those requests the same way they cover a standard API call. The announcement says they apply in the same way. Get that sentence into your own agreement, with your firm’s name on it.
The recap is short. Self-hosted Claude Managed Agents, in private beta since November 12, 2026, moves the harness, the sandboxes, the credentials and the session history into your own cloud or building. The model still answers from a provider you choose, the model bill still travels with each request, and no price for the service has been published.
Put the 3 questions to your account team this month and keep the answers with the contract file, dated. No hands-on test was run for this piece, and no price was available to check, so the checked figure to carry into your budget meeting is the absent 1: until a price arrives in writing, this line in the plan stays unpriced, and the rollout stays on hold until the price lands.