MiniMax open-sources parsar-core, a self-hosted OpenAI Agents API server
MiniMax released parsar-core, an open-source, self-hosted implementation of the OpenAI Agents API under the MIT License. It lets developers run Codex, Claude Code, or MiniMax Code with their own models via the official OpenAI SDK. The source release was published on 2026-09-21.
China context
- Original name
- 稀宇科技
- Outside China
- Open weights · github.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers can integrate an OpenAI-compatible Agents API into their own infrastructure using the official SDK, with the flexibility to swap models and sandbox providers.
- For investors
- MiniMax's open-source release may accelerate adoption of its agent technology and create an ecosystem around its tooling, potentially increasing demand for its commercial offerings.
Translated from Chinese. Quotes and facts link to the original sources.
MiniMax published parsar-core on GitHub, an open-source server that implements the OpenAI Agents API for self-hosting. It supports running Codex, Claude Code, or MiniMax Code with user-configured models through the official OpenAI SDK or plain HTTP. The system uses pluggable Sandbox Providers (including microsandbox and E2B), workspace filesystems, Harnesses, and model providers. Installation is via Docker on Linux/macOS or PowerShell on Windows, with a Web UI for configuration. The project is released under the MIT License.
parsar-core exposes the same API as OpenAI's Agents API, allowing existing OpenAI SDK clients to work unchanged by setting OPENAI_BASE_URL. It decouples agent execution from model providers and sandbox backends through defined protocols, enabling replacement of individual components. The architecture separates the Agents API (/v1) for applications from the Core API (/core/v1) for operators, with durable execution state managed by Core.
Developers outside China can now self-host an OpenAI-compatible Agents API using MiniMax's open-source code, reducing dependency on OpenAI's cloud and enabling use of alternative models. This directly lowers infrastructure costs and increases deployment flexibility for teams building agent-based applications.
For enterprises, parsar-core offers a path to run agent workloads on private infrastructure with the familiar OpenAI SDK, potentially reducing API costs and addressing data residency concerns. The MIT license permits commercial use and modification without royalty obligations.
Observable next signals include the number of GitHub stars and forks, community contributions, and whether MiniMax provides official support or updates. Verification question: does parsar-core support models from providers other than those explicitly mentioned, and what are the performance characteristics compared to OpenAI's hosted Agents API?