Open Managed Agents

Open Managed Agents is an open-source, wire-compatible implementation of the Claude Managed Agents API designed to be self-hosted on Cloudflare Workers + Durable Objects or via Docker. It provides a full agent platform: durable sessions, sandboxed code execution, vault-backed credentials, MCP server hosting, and Claude Code skills — all under your control. The project is Apache 2.0 licensed and aims to offer a transparent, inspectable alternative to closed vendor platforms. Who it's for: developers and teams who want to run AI agents without per-seat fees, usage taxes, or vendor lock-in. It suits those comfortable deploying on Cloudflare's edge or containerized environments, and who need the flexibility to customize the agent loop (the harness) or inspect every state transition. How it works: the platform manages infrastructure (sandbox, event log, tools, crash recovery) while you write the brain (event handling, model calling). Sessions are backed by Durable Objects with embedded SQLite for strong consistency, sandboxes run in Cloudflare Containers, and blobs are stored in R2. The HTTP API mirrors Claude Managed Agents exactly — same /v1/agents, /v1/sessions, and SSE event streams — so migrating is as simple as swapping the base URL. The latest blog posts detail the architecture of Durable Objects, R2, and the brain/sandbox split. What makes it different: every resource definition, state machine transition, and retry rule is in readable source code. The harness layer is explicit and replaceable — you can swap in custom context engineering, caching, or tool delivery without touching the platform. It's fully self-hostable, with a five-line quickstart via Docker Compose. Positioned as the open-source foundation for Claude Tag-style agents, enabling self-hosted Claude Tag alternatives.