OMA: a self-hosted TypeScript agent runtime and its website

Project

OMA open-source framework and website

The situation

When a company wires agents into its business, critical actions need someone to approve them and every run needs to be verifiable afterwards; keys, models and the runtime environment have to stay in the company's own hands.

Solution

OMA was released on 2026-04-01 under the MIT license and is built and maintained by YuanASI. Its premise: ownership, approval and audit of agents belong to the organization that runs them.

Durable approvals: plans, task dispatch and tool calls can all be suspended to wait for a human decision; each approval request is bound to the content the approver actually saw, and after a process restart the run resumes from that content.

Verifiable journal: it records what the model saw, every tool call and its result, so a run can be checked byte by byte offline afterwards.

Governance floor: declare which roles must actually execute, and at the end of a run execution receipts decide whether that was met.

Runs in your own environment: no telemetry, no hosted control plane; more than ten built-in model adapters, with local models connected through baseURL.

The website open-multi-agent.com carries docs and examples in Chinese and English, with its source public. Third-party projects publicly using OMA include temodar-agent and Engram; in the AI office-space platform delivered by YuanASI, the product-recommendation step is also orchestrated by OMA.

Sources

• OMA website：https://open-multi-agent.com ↗
• GitHub project：https://github.com/open-multi-agent/open-multi-agent ↗