OMA Open Multi-Agent Runtime

What does this agent do, and when should you use it?

OMA is an MIT-licensed, self-hosted TypeScript agent runtime delivered primarily through the @open-multi-agent/core library, without an OMA account or hosted control plane. It exposes runAgent(), runTasks(), and runTeam() for single-agent execution, explicit pipelines, and coordinator-planned multi-agent task graphs. The runtime can call hosted providers, local model servers, OpenAI-compatible endpoints, Vercel AI SDK providers, and external coding agents connected through process or ACP backends. Plans, task dispatches, and consequential tool calls can suspend for durable approval and later resume from checkpointed state. A journal can capture model-visible blocks, tool calls and results, and context rewrites, while verifyRun() checks their provenance offline and byte for byte. The repository also includes an offline Run Viewer, an optional @open-multi-agent/otel adapter, and the create-oma-app scaffolder, making it most relevant to teams that need to own their model credentials, deployment boundary, approvals, and audit evidence.

Developers create agents or teams with OpenMultiAgent, then use runAgent() for one agent, runTasks() for a predefined pipeline, or runTeam() to turn a goal into a task graph with assignees and dependencies. In runTeam(), one model call creates task specifications, a deterministic scheduler executes them, and a second model call produces the final response from completed task outputs; the coordinator is not consulted during execution. Configured agents can invoke tools, while plan, dispatch, or tool-call gates may return suspend. Each durable approval request is stored beside its checkpoint and bound by SHA-256 to exactly what the reviewer saw. Decisions are atomic and first-wins; altered requests and stores lacking compare-and-set fail closed. With a journal backend attached, OMA records model-visible blocks, tool calls, results, and context rewrites; verifyRun() cold-reads the record and checks whether each block can be reproduced byte for byte from its named source event. Results can expose task statuses, assignees, dependencies, token usage, pending approvals, governance receipts, and traces for the offline Run Viewer.

1. A backend team that requires human authorization before file writes or shell commands can suspend consequential calls and resume them after a service restart.
2. A regulated organization can declare required roles and independent review, then assess the actual role order and dependency edges from an execution receipt.
3. A team operating inside a restricted network can connect Ollama, vLLM, llama-server, or another OpenAI-compatible endpoint and enforce an offline or allowlist egress policy.
4. TypeScript developers prototyping pull-request review, security analysis, or teaching workflows can select a starter and run the deterministic no-key demo through create-oma-app.
5. Automation developers with a known workflow can use runTasks(), while teams that need runtime decomposition can let runTeam() construct and schedule the graph.
6. Operations or compliance teams that need inspectable context lineage can retain run journals and use verifyRun() for offline, byte-level verification.

How do you install or deploy this agent?

OMA requires Node.js 20 or newer. Production deployments should use a currently maintained Node.js LTS release. Node.js 20 is retained only as a migration window and is scheduled for removal in the next major release, no earlier than 2026-10-31.

Scaffold a starter application:

npm create oma-app@latest my-oma

In an interactive terminal, the scaffolder selects a starter and runtime, installs dependencies, and launches a deterministic local demo. That demo needs no API key and makes no model request.

To add OMA to an existing backend:

npm install @open-multi-agent/core

For the documented OpenAI example, provide credentials:

export OPENAI_API_KEY="your-api-key"

Other hosted providers, local servers, OpenAI-compatible endpoints, and AI SDK providers require their corresponding credentials or baseURL configuration.

How do you use this agent?

This example creates a team that suspends consequential tool calls and stores its checkpoint in a local file:

import { FileStore, OpenMultiAgent } from '@open-multi-agent/core'

const oma = new OpenMultiAgent({
  defaultProvider: 'openai',
  defaultModel: 'gpt-5.4',
  onToolCall: ({ consequential }) => (consequential ? { action: 'suspend' } : { action: 'allow' }),
})

const team = oma.createTeam('ops', {
  name: 'ops',
  agents: [{ name: 'operator', systemPrompt: 'Reconcile overdue invoices.', toolPreset: 'readwrite' }],
})

const result = await oma.runTeam(team, 'Find overdue invoices and draft the reminders.', {
  checkpoint: { store: new FileStore('./.oma/run.json') },
})

When result.status?.code is suspended, the host application must present and decide the entries in result.pendingApprovals; each entry is bound to the exact content shown to the reviewer. Use runTasks() when the graph is already known, or runAgent() for a single agent. When journaling is enabled, use verifyRun() to validate the saved run offline.

What are this agent's strengths and limitations?

How does this agent compare with similar options?

Key facts side by side with the most closely related agents.

How does FollowAgents rate this agent?

Recent releases

Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.

1. v1.23.0 2026-10-29
   Release notes excerpt

   Breaking changes This release ships as a minor version, so a caller on a ^1.x range receives it without an explicit upgrade step. A project that pins zod below 3.25, code that uses object-only members of taskProfileSchema or gatePolicySchema, and a coordinator plan over the new size caps can stop working after the upgrade. The individual entries are under Changed below. …

   View on GitHub ↗
2. v1.22.0 2026-10-23
   Release notes excerpt

   Breaking changes This release ships as a minor version, so a caller on a ^1.x range receives it without an explicit upgrade step. A caller that passes malformed task metadata or disables thinking on MiniMax-M3.1-Flash still installs, then fails when the code runs. The individual entries are under Changed and Fixed below. …

   View on GitHub ↗
3. v1.21.3 2026-10-16
   Release notes excerpt

   Fixed - Symlink loops and chains longer than the kernel's link limit now fail the sandbox check with a Too many levels of symbolic links error. Earlier versions also failed these writes, with the filesystem error instead. - A symlink target that ends in a separator or . must now name an existing directory, as the kernel requires. On macOS, earlier versions resolved such a link to the file of that name, so file_write overwrote it. Security …

   View on GitHub ↗

FAQ