What is Open Multi-Agent (OMA)?
Open Multi-Agent, or OMA, is a TypeScript library for running one agent, a fixed pipeline of tasks, or a team that plans its own task graph from a goal. In team mode one model call turns the goal into tasks with assignees and dependencies, a deterministic scheduler runs them, and a second call writes the answer; the coordinator is not consulted mid-run, and the finished run can be read back as data.
Its selling points are governance features: durable approvals that pause before a plan, task or tool call and survive a process restart; a run journal of model inputs, tool calls and results with an offline consistency check; declared governance roles a run can be checked against; and an egress policy that can keep model traffic offline or on an allowlist. It has adapters for Anthropic, OpenAI, Azure, Bedrock, Gemini, Grok, Copilot, DeepSeek and others, plus Ollama and vLLM, and makes no telemetry or phone-home requests.
The project launched on 31 March 2026 under MIT and is maintained by YuanASI, a Shenzhen company that also builds systems on it for clients. It has about 7,000 GitHub stars; @open-multi-agent/core is at 1.23.0 (29 October 2026) with about 840 npm downloads in the last week of October.
Is Open Multi-Agent (OMA) worth it?
OMA's design reads like it was written by someone who has had to explain an agent's actions to an auditor. Approvals are durable, so a run paused for sign-off can resume from the reviewed content after a restart; your application supplies the approval screen. The run journal records what the model saw and did, and verifyRun() checks it offline, with the documentation careful to say this proves consistency, not that nobody edited it. The governance check is similarly honest: a run can succeed and still fail its declared role requirements, and the caller has to look.
The egress policy is the feature most frameworks lack. Set it to offline or an allowlist, and built-in adapters refuse to connect elsewhere; a child policy can only tighten its parent, and a transport that cannot be enforced fails closed. The README notes that process and ACP backends, which put Claude Code, Gemini CLI or Codex on the same task graph, sit outside that policy. There is no OMA backend or account, and the library makes no analytics, licence or update requests.
Adoption is the open question. The repository shows about 7,000 stars and an unusually high 2,440 forks, yet the core npm package recorded about 840 downloads in the last week of October, and the named production users fit in a short list. Releases are frequent and not always gentle: 1.23.0 shipped as a minor version but made zod a required peer dependency, changed two schema types and capped coordinator plans at 200 tasks, any of which can break a caller on a ^1.x range. Node 20 support is due to be removed in the next major.
Against Mastra or LangGraph, OMA is narrower and far less proven, with a smaller ecosystem of integrations and examples. Its case is specific: a TypeScript team that needs approvals, an inspectable record and enforced egress more than it needs breadth.
Where it earns its place
A backend engineer at a 100-person fintech
Add an agent step that needs sign-off before it touches customer records
Wrap the tool call in a durable approval with persistent storage, wire the approval request into the existing admin app, and keep the run journal for compliance review. MIT; the cost is model tokens and the engineering to build the approval screen.
A security lead at a company with a no-external-AI rule for some data
Prove that an agent workflow cannot call an outside model
Run OMA against local Ollama or vLLM with the egress policy set to offline, so built-in adapters refuse other hosts. Process and ACP backends are outside the policy and need separate controls.
A developer comparing multi-agent frameworks for a Node service
See how a planned task graph behaves before writing code against it
Run npm create oma-app for the no-key scripted demo and inspect the offline Run Viewer's task DAG and span waterfall. Weigh the small install base and minor-version breaking changes before committing.
Before you commit
- Setup
- npm create oma-app@latest for a starter with a no-key demo, or npm install @open-multi-agent/core into an existing backend. Node 20 or newer, with a maintained LTS recommended; zod ^3.25 or ^4.6.6 is a required peer from 1.23.0.
- Integrations
- Anthropic, OpenAI, Azure OpenAI, Bedrock, Gemini, Grok, Copilot, DeepSeek, Doubao, Hunyuan, MiniMax, MiMo and Qiniu adapters; Ollama, vLLM and llama-server by baseURL; Vercel AI SDK providers; MCP; Claude Code, Gemini CLI and Codex as external agents; optional OpenTelemetry package.
- Data and privacy
- A library with no telemetry, no OMA backend and no phone-home requests; traces and journals stay in your own store. Prompts go to whichever provider you configure, within the egress policy. MIT licence.
- Cost at scale
- No paid edition; cost is model tokens and your own hosting. Team mode adds two coordinator calls per run on top of the agents' own work.
- Limits
- About nine open issues; plans capped at 200 tasks from 1.23.0; minor releases can break ^1.x callers; Node 20 support ends in the next major; small production user base.
What’s good
- Durable approvals that survive a process restart
- Egress policy fails closed and can only tighten
- No telemetry, account or hosted control plane
- No-key local demo drives the real scheduler
What’s not
- About 840 weekly npm downloads against 7,000 stars
- Minor releases have carried breaking changes
- Journal check proves consistency, not tamper-resistance
- Smaller ecosystem than Mastra or LangGraph
Best for
TypeScript teams that must show who approved each agent action and where model traffic went, and can live with a young, single-company project.
Open Multi-Agent (OMA) alternatives
Open Multi-Agent (OMA): common questions
- Is Open Multi-Agent (OMA) worth it?
- OMA's approvals, journal and egress controls are well thought through, but a few hundred weekly installs behind 7,000 stars says it has more admirers than users. We rate it 3.6 out of 5.
- How much does Open Multi-Agent (OMA) cost?
- Free and open source (MIT). No hosted service; the maintainer sells custom builds on request.
- What are the downsides of Open Multi-Agent (OMA)?
- About 840 weekly npm downloads against 7,000 stars; minor releases have carried breaking changes; journal check proves consistency, not tamper-resistance; smaller ecosystem than Mastra or LangGraph.
- Who is Open Multi-Agent (OMA) best for?
- TypeScript teams that must show who approved each agent action and where model traffic went, and can live with a young, single-company project.
- How much does Open Multi-Agent (OMA) cost for a team or in production?
- No paid edition; cost is model tokens and your own hosting. Team mode adds two coordinator calls per run on top of the agents' own work.
- Is Open Multi-Agent (OMA) safe for company data?
- A library with no telemetry, no OMA backend and no phone-home requests; traces and journals stay in your own store. Prompts go to whichever provider you configure, within the egress policy. MIT licence.
- What are the best Open Multi-Agent (OMA) alternatives?
- The closest alternatives we have reviewed are Mastra, LangGraph, CrewAI and OpenAI Agents SDK.
Answers checked Oct 31, 2026.