XNews

MiniMax Open-Sources OpenAgentCore Self-Hosted OpenAI Agents API

MiniMax Open-Sources OpenAgentCore Self-Hosted OpenAI Agents API

MiniMax released OpenAgentCore, an MIT-licensed, self-hosted implementation of the OpenAI Agents API that runs Codex, Claude Code or MiniMax Code on user machines with user-configured models through the official OpenAI SDK. The system defines ten protocol boundaries decoupling harness, model provider, sandbox and filesystem, and persists agent state across machine failures via NFS files, VM checkpoints and native session history.

Event version 1.

Key insights

Standard API, swappable internals

OpenAgentCore exposes one standard API externally while harness, model, sandbox and filesystem are all protocolized internally. Ten defined boundaries decouple the modules, so adding a new harness, model provider or sandbox requires only implementing the corresponding adapter.

Native harness execution

The project does not build its own agent loop; each harness runs natively and retains control of permissions, scheduling and context. Core acts as the control plane, assembling sandbox, model endpoint and configuration and passing them through.

Three-layer state persistence

Files are stored on NFS, memory and process state are captured as VM checkpoints that can restart on the original or another node in seconds, and harness-native session history is stored alongside files so context survives a full environment rebuild.

SDK compatibility and deployment options

The official OpenAI SDK or raw HTTP can point at a Core /v1 endpoint, and each session selects a harness such as codex while models route through the configured provider. Sandboxes support microsandbox, E2B pause, Docker and local Linux, macOS or Windows.

Source posts