Self-hosted Managed Agents: what still reaches Anthropic, per inference provider
The question
"What do Anthropic's own docs say that the self-hosted Claude Managed Agents private beta (agent server, database, sandboxes, credentials and session history in the customer environment, selectable model provider) actually keeps on Anthropic's side, and which data-residency objections does it resolve when the model provider is Anthropic's API versus Bedrock, Google or an internal gateway?"
Context: the only vault evidence was a newsletter blurb. This brief replaces it with the primary docs, which turned out to be publicly readable even though the software itself is gated.
Evidence labels used below: [P] = read in Anthropic's own docs or announcement on 2026-11-19. [S] = secondary source only. [I] = my inference, not stated by any doc.
What we already know (from the vault)
- [[2026-11-13-alphasignal-claude-meetings-self-hosted-managed-agents-anthropic-ipo]] is the only prior coverage: one blurb saying server, database, sandboxes, credentials and session history stay in the customer environment, with a selectable model provider. [S]
- [[2026-07-02-claude-sandboxes-security-pricing-vs-rdco-substrate]] covered the middle option, self-hosted sandboxes: tools run on your compute, but Anthropic still runs the agent loop and the session store, and tool inputs and outputs flow through Anthropic's control plane.
- [[2026-10-19-headless-cron-vs-managed-agents-metering]] established that hosted Managed Agents is metered against a Console (API) organization, separate from subscription usage. It did not cover self-hosted.
- [[2026-07-09-stratechery-update-muse-grok-karp]] logged "why would we give Anthropic access to our workflows?" as a live regulated-client objection, and recorded a secondary claim that Fable-class models carry a mandatory 30-day retention even for zero-retention customers. That claim was not re-verified here and matters for the Claude API row below.
What the web says
- The product and its gating. Announced 2026-11-12 in "Operate Claude Managed Agents inside your security perimeter" [P]. Private beta; access to the download is through the Anthropic account team [P]. The docs pages themselves are public (they carry a "request access" note but render in full) [P]. No pricing is stated in the announcement or in the three docs pages read [P for the absence; S for "pricing not disclosed"].
- Who hosts what. The docs give a three-row table: cloud sandboxes (Anthropic hosts sandboxes, service and model), self-hosted sandboxes (you host sandboxes only), self-hosted Managed Agents (you host sandboxes and service; model is "Anthropic, or a provider that you choose"). "The service" means the API, the agent loop, and storage for agents, sessions and session history. Components you run:
cma-server, PostgreSQL 16+, workers, a console page on your server, and an admin CLI. You install each release yourself. [P] (self-hosted overview) - The headline sentence. "The server connects to Anthropic only when that provider is the Claude API. Anthropic has no remote access to the server or its log." Anthropic issues and holds none of the credentials the server accepts. The server's full connection list is: inference provider, its database, credential-issuing services, a cloud model-listing service, MCP servers and Git hosts you allow, an OpenTelemetry collector, identity providers, a secret manager or KMS, and anything else you configure. No license-check, update-check or telemetry endpoint appears in that list. [P] (data and access)
- What every model call carries, to whichever provider. "The conversation so far, including prompts, tool inputs, and tool results." So the content of the work leaves the server on every turn; only the destination changes. [P]
- What is extra when the provider is the Claude API. Three headers with the session, agent and environment IDs (IDs hold no name or content but encode creation time; "No setting turns them off"); a
User-Agentnaming the server software and version; a fixedmetadata.user_idlabel ofcma-server; and a credential validity check at startup and every 60 seconds that runs no inference. The docs say plainly that Anthropic can use these to attribute and count usage. [P] - What is different under other providers. The three ID headers are never sent to any provider other than the Claude API (or a proxy named in
ANTHROPIC_BASE_URL). Thecma-serverlabel is sent to Microsoft Foundry and to gateways that take Claude API format requests; it is never sent to Amazon Bedrock, Claude Models on Google Cloud, or gateways using another request format. Under every provider exceptanthropic, the server needs no Claude API key. [P] (inference provider) - Retention and ZDR. Session history stays in your database until you delete it; the server keeps no log files, traces or metrics. Model inputs and outputs "fall under your organization's agreement with the provider that serves the model." With the Claude API, an existing ZDR arrangement covers these calls "in the same way as other Messages API calls." The docs also explain why hosted Managed Agents is listed as not covered by ZDR: Anthropic stores the transcripts when it runs the service, including under self-hosted sandboxes. [P]
- Region defaults worth knowing. The Bedrock sample defaults to a
us.anthropic.prefix, "which names a cross-region inference profile"; another profile family is set withmodel_id_prefix. The Google option requires aregion(a specific location orglobal) and bills to the namedproject_id. [P] - Responsibilities shift to you. Everything the self-hosted sandboxes security page assigns to Anthropic becomes the customer's job, "including securing the server, hosting memory stores, and detecting unusual usage." One deployment serves one tenant at one trust level; separate tenants need separate deployments. Cloud environments do not exist on a self-hosted server. Web search and web fetch, where offered, are run by the inference provider and are off by default. [P]
What crosses to Anthropic, by configuration
| Item | Claude API | Amazon Bedrock | Google Cloud | Microsoft Foundry | Own gateway |
|---|---|---|---|---|---|
| Server connects to Anthropic | Yes [P] | No [P] | No [P] | No [P] | No, unless the gateway forwards there [P] |
| Prompts, tool inputs, tool results | To Anthropic [P] | To AWS [P] | To Google [P] | To the Foundry resource [P] | To the gateway [P] |
| Session, agent, environment ID headers | Sent, cannot be disabled [P] | Never [P] | Never [P] | Never [P] | Only via ANTHROPIC_BASE_URL proxy [P] |
cma-server usage label |
Sent [P] | Never [P] | Never [P] | Sent [P] | Sent if Claude API format [P] |
| Credential check every 60s | Yes [P] | Not described | Not described | Not described | Not described |
| Session history, agents, vault secrets at rest | Customer [P] | Customer [P] | Customer [P] | Customer [P] | Customer [P] |
| Anthropic remote access to server or logs | None [P] | None [P] | None [P] | None [P] | None [P] |
| Retention terms for model I/O | Your Anthropic agreement, ZDR if you have it [P] | Your AWS agreement [P] | Your Google agreement [P] | Your Microsoft agreement [P] | Whoever the gateway forwards to [I] |
Convergences and contradictions
- Convergence: the newsletter blurb was accurate as far as it went. The docs confirm every component it listed stays in the customer environment.
- Where the blurb undersells and oversells. It undersells: the docs go further than "stays in your environment" and state no Anthropic remote access and no Anthropic connection at all off the Claude API. It oversells by omission: "session history stays with you" is true at rest, but the same history is re-sent as conversation context on every model call. With the Claude API as provider, the content still reaches Anthropic in transit and is governed by retention terms, not by architecture.
- Correction to the vault's July framing. The July brief described self-hosted sandboxes as a "clean data-flow boundary." Against the new three-row table, that option leaves transcripts stored at Anthropic and outside ZDR. Only the full self-hosted service moves the transcript store.
Synthesis for RDCO
The honest one-line answer for a regulated client is: self-hosted Managed Agents moves the system of record, not the inference. Anthropic's side holds no session store, no credentials, no control plane and no remote access in any configuration. What Anthropic still sees depends entirely on the inference setting. With the Claude API it sees the full conversation on every turn plus non-content identifiers it uses to count usage; the residency story there is contractual (ZDR, and whatever model-specific retention terms apply). With Bedrock or Google, the docs say the server never connects to Anthropic and sends no Anthropic usage identifiers at all, so the conversation is governed by the client's existing cloud agreement. That is the configuration that answers "nothing goes to Anthropic" in a form a security reviewer can check against an egress log, which the docs explicitly invite.
So the objections sort into three groups. Resolved in every configuration: transcripts stored at a vendor, vendor-held credentials, vendor-operated control plane, vendor access to logs, and hosted Managed Agents falling outside ZDR. Resolved only off the Claude API: prompts and tool results transiting Anthropic, and Anthropic-side usage attribution by session. Not resolved by this product at all: the model still runs somewhere outside the client's own hardware, so in-region processing depends on the provider's region settings (the Bedrock default is a US cross-region profile, a detail that would fail an EU or single-region requirement if left untouched), and the client now owns server hardening, tenant separation by deployment, anomaly detection and upgrades. Microsoft Foundry sits in between: no Anthropic connection from the server, but the cma-server label is sent and the docs note Anthropic can use it "where a call reaches Anthropic." The docs read here do not say whether Foundry-served calls reach Anthropic, so that row should not be presented as equivalent to Bedrock without checking Microsoft's and Anthropic's Foundry terms. [I]
For phData positioning this is a cleaner story than anything available before 2026-11-12: a client already approved for Claude on Bedrock or Google can run the same Managed Agents API and SDK with no new data processor added, and the agent definitions work unchanged across providers because the server translates model IDs. The costs to state up front are that it is a private beta with account-team gating, undisclosed pricing, a feature gap versus hosted (a "differences" page exists and was not read in full; cloud environments are confirmed absent), and a real operations burden. It is a fit for a client with a platform team, not a shortcut.
For the cert topic, the three-row hosting table is the thing to memorize, with one trap: self-hosted sandboxes do not change where transcripts live or ZDR coverage; only self-hosted Managed Agents does, and choosing a non-Anthropic provider is what removes the Anthropic connection. Note this beta postdates most study material, so an exam item written earlier may treat "self-hosted harness on your infrastructure with a ZDR org" as the only answer.
Why this is in the vault
It gives the phData regulated-client conversation a doc-backed, per-provider answer to "what does Anthropic still see?" for agent deployments, and it updates the self-hosted vs Anthropic-hosted distinction used in cert study now that a third hosting option exists.
Open follow-ups
- What does the "Differences from the hosted service" page list as missing on self-hosted, overall and per inference provider (memory, skills, web tools, multi-agent features)?
- How is self-hosted Managed Agents priced and licensed: does the Software License Agreement impose a fee, usage reporting or audit rights when inference runs entirely on Bedrock or Google, where Anthropic receives no usage identifiers?
- When Claude is served through Microsoft Foundry, do requests reach Anthropic-operated infrastructure, and under whose retention terms?
- Does any current Claude model carry mandatory retention that overrides ZDR on the Claude API path, and does the same condition apply to that model on Bedrock or Google?
- Which Bedrock inference-profile families and Google locations keep processing in a single region for current Claude models, so the
model_id_prefixandregionsettings can satisfy an in-region requirement?
Related
- [[2026-11-13-alphasignal-claude-meetings-self-hosted-managed-agents-anthropic-ipo]] - the secondary blurb this brief replaces with primary docs
- [[2026-07-02-claude-sandboxes-security-pricing-vs-rdco-substrate]] - the middle hosting option (self-hosted sandboxes) this extends
- [[2026-10-19-headless-cron-vs-managed-agents-metering]] - metering of hosted Managed Agents against a Console organization
- [[2026-07-09-stratechery-update-muse-grok-karp]] - the client objection about Anthropic seeing workflows, and the unverified retention claim
- [[2026-10-20-ccdv-f-quiz-1]] - cert practice material touching the self-hosted vs hosted and ZDR distinction
Sources
- Vault: [[2026-11-13-alphasignal-claude-meetings-self-hosted-managed-agents-anthropic-ipo]]
- Vault: [[2026-07-02-claude-sandboxes-security-pricing-vs-rdco-substrate]]
- Vault: [[2026-10-19-headless-cron-vs-managed-agents-metering]]
- Vault: [[2026-07-09-stratechery-update-muse-grok-karp]]
- Vault: [[2026-10-20-ccdv-f-quiz-1]]
- Web, primary, read in full: Claude Platform Docs, "Self-hosted Claude Managed Agents" - https://platform.claude.com/docs/en/managed-agents/self-hosted
- Web, primary, read in full: Claude Platform Docs, "Data and access in self-hosted Claude Managed Agents" - https://platform.claude.com/docs/en/managed-agents/self-hosted-data-and-access
- Web, primary, read in full: Claude Platform Docs, "Inference provider" - https://platform.claude.com/docs/en/managed-agents/self-hosted/inference-provider
- Web, primary, read via summarizer (quotes are from the summary, not checked against the page): "Operate Claude Managed Agents inside your security perimeter," 2026-11-12 - https://claude.com/resources/articles/self-hosted-claude-managed-agents
- Web, primary, search snippet only, not read: "Differences from the hosted service" - https://platform.claude.com/docs/en/managed-agents/self-hosted/differences
- Web, primary, not read: Software License Agreement, Claude Managed Agents Server - https://www.anthropic.com/legal/self-hosted-managed-agents-license
- Web, secondary, search snippets only: aiweekly.co, thenewstack.io and vktr.com coverage of the 2026-11-12 announcement (used only for "pricing not disclosed" and the date)
- Research caps: QMD 2 of 5, WebSearch 2 of 3, WebFetch 4 of 3 (one over, for the inference-provider page). No paywalls.