open-multi-agent/open-multi-agent

What it solves

Open Multi-Agent (OMA) provides a way to orchestrate multiple AI agents to achieve a complex goal without requiring the developer to manually hard-code every step of the workflow. It allows for dynamic planning where the task graph emerges at runtime based on the goal, while still maintaining deterministic control, inspectability, and the ability to run in secure, air-gapped, or local environments.

How it works

OMA uses a coordinator to decompose a high-level request into a task Directed Acyclic Graph (DAG) at runtime. This plan is then executed by a team of agents. The framework provides tools for previewing, approving, and replaying plans, as well as multi-agent consensus to ensure reliability. It is built for TypeScript backends and can be integrated into Node.js applications, supporting a mix of cloud-based models (like GPT and Claude) and local open-source models.

Who it’s for

TypeScript developers and teams building AI agentic workflows that require dynamic runtime planning, strict execution controls, and the the ability to run on their own infrastructure (local or offline).

Highlights

• Dynamic Planning: Decomposes goals into task DAGs at runtime rather than using hand-wired graphs.
• Self-Hosted: Runs in your own environment, supporting offline and air-gapped setups with default-deny tools and auto-redacted secrets.
• Model Agnostic: Mixes cloud and local models on a single team, including a fallback parser for local models that handle tool calls as text.
• Observability: Includes an offline Run Viewer for inspecting and replaying traces, with optional OpenTelemetry integration for enterprise monitoring.
• Deterministic Control: Features plan preview/approval, plan freezing/replay, and multi-agent consensus mechanisms.