MiniMax Open-Sources Self-Hosted OpenAI Agents API "OpenAgentCore"
MiniMax has open-sourced a self-hosted OpenAI Agents API called "OpenAgentCore". With the official OpenAI SDK or plain HTTP, you only need to point the address to the /v1 endpoint of your self-built Core to make calls. Each Session can optionally use a harness (such as codex), natively running Codex, Claude Code, and MiniMax Code within the session, with models going through a self-configured provider. The Sandbox Provider, workspace filesystem, Harness, and model provider are all replaceable according to established protocols, supporting microsandbox, E2B, Docker, and local Linux/macOS/Windows.
Why read this
AI score 74
The original text explains the interface boundaries and replaceable scope of this self-hosted Agent Infra, so readers can judge the migration cost of existing Agent workflows accordingly.
Topics
View original post[1/n] In the past, people called APIs as Tokens as a Service, but now it is gradually becoming Agents as a Service. However, there is a very practical problem in building this layer: models and harnesses basically iterate every few months now, and if you tie your business directly to a certain implementation, switching later will be extremely painful. So the underlying Agent Infra needs to keep its external interfaces as stable as possible, while internally being able to swap in the most frontier components at any time. This is the original intention behind our work on OpenAgentCore, an open-source Agent Infra: externally it provides only one standard API, while internally everything—whether harness, model, sandbox, or Filesystem—is protocolized and can be freely replaced, so you won't be locked into any single vendor. https://t.co/V0xdSK4UPr
View original
[1/n] 以前大家调 API 是 Tokens as a Service，现在慢慢变成了 Agents as a Service。 但做这层存在一个很现实的问题：现在模型和harness基本几个月就迭代一轮，如果把业务直接绑死在某套实现上，后面换起来会非常痛苦。所以底层的 Agent Infra 尽可能需要做到对外的接口足够稳定，但是对内得能够随时换上最 frontier 的组件。 这就是我们做 OpenAgentCore 的初衷，一个开源的 Agent Infra：对外只提供一套标准的 API，对内不管是 harness、model、sandbox 还是 Filesystem 全部协议化，都可以随意更换，不会被绑在某一家上。 https://t.co/V0xdSK4UPr
Body · AI translation
MiniMax has open-sourced a self-hosted OpenAI Agents API, "OpenAgentCore": one Core, multiple Harnesses; the official SDK only needs OPENAI_BASE_URL changed, and Codex / Claude Code / MiniMax Code run natively within a Session, with models and sandboxes swappable by protocol
Open-source address: https://t.co/EMQtUxhyBN
Main problems it aims to solve 1. Same API: the official OpenAI SDK or raw HTTP, pointed at your Core /v1. 2. Your Agent + your model: each Session selects a harness (e.g. extra_body.agent.x_agents_core.harness = "codex"), and the model goes through the provider you configure. 3. Compute can sleep, workspaces can persist: on microsandbox, idle compute is released and files are restored with independent storage on compatible nodes; E2B supports in-place pause; Docker and local Linux/macOS/Windows are also supported. 4. Boundaries are replaceable: Sandbox Provider, workspace filesystem, Harness, and model provider all follow established protocols (see AGENTS.md).
Project architecture · Core + PG: authentication, frozen configuration, scheduling Turns, cancellation/interaction, persisting execution facts · Sandbox Provider / Node: requesting and reclaiming compute; nodes run Docker or microsandbox · Workspace FS: persistent environment storage (separate from the compute lifecycle) · Runtime: prepares the environment and capability snapshot, runs the Session Executor, reports events and receipts · Harness adapter: calls upstream SDKs/protocols, translates events; the native agent loop and context belong to the Harness · Model provider: Responses / Anthropic / chat_completions and other protocols · Web: operations console (via /core/v1)