Author: Anthropic URL: https://claude.com/resources/articles/self-hosted-claude-managed-agents Published: November 12, 2026 Product: Claude Platform, Claude Managed Agents
Self-hosted Claude Managed Agents (private beta) lets enterprises run the full Managed Agents service inside their own cloud or on-premise environment — harness, sandboxes, credentials, and session history — while still using the same developer APIs as the hosted service.
What Changed
Managed Agents launched as a hosted service (Anthropic runs the infrastructure). Then added self-hosted sandboxes for teams who needed their code in their own environment. Now: full self-hosting for teams that want everything — harness, database, workers — in their own perimeter.
What Self-Hosting Covers
- Harness + sandboxes: both run in your environment
- Credentials: stored and managed by you, never sent to Anthropic
- Session history: stored in your own database
- Model routing: requests go to Anthropic directly, Amazon Bedrock, Google’s Agent Platform, Microsoft Foundry, or your own gateway
- Developer APIs: identical to the hosted service — no code changes required to switch hosting modes
- ZDR compatibility: if zero data retention is in place with Anthropic, it applies to model requests as with any other Messages API call
Four Agent Security Principles
These apply to all Managed Agents deployment modes:
- Brain and hands are split — The harness (brain) calls Claude; the sandbox (hands) is where agent code runs. These are kept separate. Credentials and session history stay in the harness.
- Credentials stay in the harness — Tool credentials are never exposed to the sandbox. The harness makes tool calls on behalf of the agent.
- Resources are scoped — Sandboxes have only the access they need; each is isolated by default.
- Actions are logged — Every tool call is logged for review.
Who This Is For
Teams running agents on critical internal systems — fixing builds, testing releases, triaging alerts, working bug backlogs — where the entire service needs to stay inside the enterprise perimeter, not just the sandbox.
Implementation
Tool/Service: Claude Managed Agents (self-hosted mode) Setup: Private beta — apply via Anthropic. Docs: https://platform.claude.com/docs/en/managed-agents/self-hosted Cost: Contact Anthropic (enterprise pricing) Integration notes: Works with Bedrock, GCP Agent Platform, Azure Foundry, or bring-your-own gateway. Session history requires operating your own database.
Key Takeaways
- Self-hosted Managed Agents is in private beta (Nov 12, 2026)
- Full self-hosting: harness + sandboxes + credentials + session history all stay in your environment
- Same APIs as hosted mode — no code changes required
- Model requests can be routed to your cloud provider or gateway instead of Anthropic directly
- ZDR agreements apply automatically
- Designed around four security principles (split brain/hands, harness-held credentials, scoped resources, full logging)
Related
Source: claude-blog-self-hosted-managed-agents-2026-11-12.md