open-multi-agent-open-multi-agent/packages/core/README.md

Open Multi-Agent

From a goal to a task DAG, automatically.
TypeScript-native multi-agent orchestration.

English · 中文

open-multi-agent is a multi-agent orchestration framework for TypeScript backends. Give it a goal; a coordinator agent decomposes it into a task DAG, parallelizes independents, and synthesizes the result. Drops into any Node.js backend.

Your engineers describe the goal, not the graph.

Graph-first frameworks make you enumerate every node and edge up front. open-multi-agent is goal-first: you describe the outcome and the coordinator builds the task DAG at runtime, so the orchestration adapts to the goal instead of being hand-wired for one.

Contents

Quick Start · Three Ways to Run · Features · Orchestration Controls · Ecosystem · Examples · How Is This Different? · Architecture · Supported Providers · Production Checklist · Documentation · Contributing

Quick Start

Requires Node.js >= 18.

The fastest way to see a multi-agent run — scaffold a project and start it in one command:

npm create oma-app@latest

The first run shows the coordinator decompose one goal into a multi-agent DAG, then opens a dashboard of the run. To add the library to an existing project instead:

npm install @open-multi-agent/core

Migrating from @jackchen_me/open-multi-agent? That package is deprecated; install @open-multi-agent/core instead.

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

// Works with any OpenAI-compatible provider. Set OPENAI_API_KEY for OpenAI, or
// set OPENAI_BASE_URL + OMA_MODEL for Groq, DeepSeek, Ollama, etc.
const model = process.env.OMA_MODEL ?? 'gpt-5.4'

// Built-in tools are opt-in (default-deny): each agent gets only the tools it
// lists in `tools` (or a `toolPreset`). List neither and the agent gets none.
const agents: AgentConfig[] = [
  { name: 'architect', model, systemPrompt: 'Design clean API contracts.', tools: ['file_write'] },
  { name: 'developer', model, systemPrompt: 'Implement runnable TypeScript.', tools: ['bash', 'file_read', 'file_write', 'file_edit'] },
  { name: 'reviewer', model, systemPrompt: 'Review correctness and security.', tools: ['file_read', 'grep'] },
]

const orchestrator = new OpenMultiAgent({
  defaultProvider: 'openai',
  defaultModel: model,
  defaultBaseURL: process.env.OPENAI_BASE_URL, // unset = OpenAI
  onProgress: (event) => console.log(event.type, event.task ?? event.agent ?? ''),
})

const team = orchestrator.createTeam('api-team', { name: 'api-team', agents, sharedMemory: true })

// Built-in filesystem tools default to a `<cwd>/.agent-workspace` sandbox.
const result = await orchestrator.runTeam(
  team,
  `Create a REST API for a todo list in ${process.cwd()}/.agent-workspace/todo-api/`,
)

console.log(result.success, result.totalTokenUsage.output_tokens)

Run an example locally

git clone https://github.com/open-multi-agent/open-multi-agent && cd open-multi-agent
npm install
export OPENAI_API_KEY=sk-...
npx tsx packages/core/examples/basics/team-collaboration.ts

Three agents collaborate on a REST API while onProgress streams the coordinator's task DAG:

agent_start coordinator
task_start design-api
task_complete design-api
task_start implement-handlers
task_start scaffold-tests         // independent tasks run in parallel
task_complete scaffold-tests
task_complete implement-handlers
task_start review-code            // unblocked after implementation
task_complete review-code
agent_complete coordinator        // synthesizes final result
Success: true
Tokens: 12847 output tokens

Local models via Ollama need no API key, see providers/ollama. For hosted providers (OPENAI_API_KEY, GEMINI_API_KEY, etc.), see Supported Providers.

Three Ways to Run

Mode	Method	When to use	Example
Single agent	runAgent()	One agent, one prompt	basics/single-agent
Auto-orchestrated team	runTeam()	Give a goal, let the coordinator plan and execute	basics/team-collaboration
Explicit pipeline	runTasks()	You define the task graph and assignments	basics/task-pipeline

For answers that need scrutiny, runConsensus() runs a proposer→judge verification loop (with an opt-in per-task verify hook). See Consensus.

Preview the coordinator's task DAG without executing it, or pin that plan and replay the same graph later without another coordinator call:

// Decompose once and review the plan
const preview = await orchestrator.runTeam(team, goal, { planOnly: true })

// Turn it into a diffable, version-controllable artifact (plain JSON)
const plan = orchestrator.createPlanArtifact(preview)

// Later: replay the exact graph (same task ids, deps, assignees), no coordinator
const result = await orchestrator.runFromPlan(team, plan)

Route orchestration phases to different models with an opt-in modelRouting policy: a flagship model plans, a cheap model runs the leaf tasks. Match by phase, agent, task role/priority, or leaf status; first match wins, and omitting it leaves model selection unchanged. See Model routing.

Features

Capability	What you get
Goal-driven coordinator	One runTeam(team, goal) call decomposes the goal into a task DAG, parallelizes independents, and synthesizes the result. Unassigned tasks are auto-scheduled — dependency-first (default), round-robin, least-busy, or capability-match.
Mix providers in one team	13 built-in providers plus any OpenAI-compatible endpoint (Ollama, vLLM, LM Studio, OpenRouter, Groq), mixed freely in one team. Local servers that emit tool calls as plain text are recovered by a fallback parser. (full list · setup)
Extended thinking / reasoning	One thinking config maps to Anthropic thinking, Gemini thinkingConfig, and OpenAI reasoning_effort; reasoning is streamed as events, with opt-in preservation across a provider switch. (cross-provider-reasoning)
Tools + MCP	6 built-in (bash, file_*, grep, glob), all opt-in (default-deny — grant via tools / toolPreset), plus delegate_to_agent handoff (cycle + depth guards), custom tools via defineTool() + Zod, stdio MCP servers via connectMCPTools(). (tool config)
Streaming + structured output	Token-by-token streaming on every adapter (per-agent during team runs via onAgentStream); Zod-validated final answer with auto-retry on parse failure. (structured-output)
Human-in-the-loop	Gate execution with onPlanReady (approve the plan before any agent runs) and onApproval (approve between task rounds), or inspect first with planOnly.
Pin and replay plans	Serialize a planOnly decomposition with createPlanArtifact, then runFromPlan replays the exact task graph without re-invoking the coordinator. (patterns/plan-replay)
Lifecycle hooks + cancellation	beforeRun rewrites the prompt, afterRun post-processes or rejects the result; pass an AbortSignal to cancel a run in flight.
Configurable coordinator	Override the coordinator's model, provider, adapter, system prompt, or tools via runTeam(team, goal, { coordinator }).
Observability	onProgress events, onTrace spans, post-run HTML dashboard rendering the executed task DAG. API keys and tokens are redacted from traces, bash output, and the dashboard. (observability guide)
Pluggable shared memory	Default in-process KV; swap in Redis / Postgres / your own backend by implementing MemoryStore. (shared memory)
Checkpoint & resume	Opt-in per-run checkpointing over any MemoryStore: snapshot on each completed task, then restore() skips finished tasks to continue after a crash or restart. Best-effort saves never take the run down. (checkpoint & resume)
Sandboxed filesystem workspace	Built-in filesystem tools are sandboxed to <cwd>/.agent-workspace by default; agents sharing the default configuration share this root. For per-agent isolation, set AgentConfig.cwd; for a different shared root, set OrchestratorConfig.defaultCwd; pass null to disable. (sandbox config)

Production controls (context strategies, task retry with backoff, loop detection, tool output truncation/compression) are covered in the Production Checklist.

Orchestration Controls

Fine-grained control over a runTeam run. All optional; defaults keep behavior unchanged.

Inject team context. Prepend the goal, roster, and this worker's role to every worker prompt — helps workers stay aligned and makes multi-step runs easier to debug. Off by default; worker prompts stay byte-identical when omitted.

await orchestrator.runTeam(team, goal, { revealCoordinator: true })

Approve before running. Inspect the coordinator's plan before any agent executes, and again between task rounds. These live on the orchestrator. Returning false aborts; remaining tasks are marked skipped.

const orchestrator = new OpenMultiAgent({
  onPlanReady: async (tasks) => tasks.length <= 10,        // gate the whole plan
  onApproval:  async (completed, next) => next.length > 0, // gate each round
})

Cancel a run. Pass an AbortSignal; aborting stops the run in flight.

const controller = new AbortController()
const run = orchestrator.runTeam(team, goal, { abortSignal: controller.signal })
// controller.abort() from elsewhere to cancel

Configure the coordinator. Give the planner its own model, adapter, or extra instructions without touching the worker agents.

await orchestrator.runTeam(team, goal, {
  coordinator: { model: 'claude-opus-4-6', instructions: 'Prefer fewer, larger tasks.' },
})

Fan-out without dependencies. For MapReduce-style parallelism, use AgentPool.runParallel() directly. See patterns/fan-out-aggregate.

Shell & CI. Use the JSON-first oma binary. See docs/cli.md.

Ecosystem

open-multi-agent launched 2026-04-01 under MIT. Known users and integrations to date:

Built with OMA

• temodar-agent (~60 stars). WordPress security analysis platform by Ali Sünbül. Uses our built-in tools (bash, file_*, grep) directly inside a Docker runtime. Confirmed production use.
• PR-Copilot. AI pull-request review assistant by kidoom. Runs an OMA review team (coordinator + scoped reviewer agents), defines repo-context tools with defineTool, and adds a custom ContextStrategy for token-aware PR-diff compression. Public code on @open-multi-agent/core.

Using open-multi-agent in production or a side project? Open a discussion and we will list it here.

Integrations

• Engram — "Git for AI memory." Syncs knowledge across agents instantly and flags conflicts. (repo)
• @agentsonar/oma — Sidecar detecting cross-run delegation cycles, repetition, and rate bursts.

Built an integration? See the integration guide for how to submit a reference or vendor example and get your product listed.

Provider community offers

Limited-time provider offers for open-multi-agent users. Listings are not paid endorsements.

• MiniMax — Use MiniMax M3 in OMA's TypeScript multi-agent workflows. OMA users get 12% off the MiniMax Token Plan until 2026-06-30. See the MiniMax setup guide.

Featured partner

For products and platforms with a deep open-multi-agent integration. See the Featured partner program for terms and how to apply.

Examples

examples/ is organized by category: basics, cookbook, patterns, providers, and integrations. See examples/README.md for the full index. (production/ is open for contributions — see the acceptance criteria.)

Real-world workflows (cookbook/)

End-to-end scenarios you can run today. Each one is a complete, opinionated workflow.

• contract-review-dag: four-task DAG for contract review with parallel branches and step-level retry on failure.
• meeting-summarizer: three specialised agents fan out on a transcript, an aggregator merges them into one Markdown report with action items and sentiment.
• competitive-monitoring: three parallel source agents extract claims from feeds; an aggregator cross-checks them and flags contradictions.
• translation-backtranslation: translate EN to target with one provider, back-translate with another, flag semantic drift.
• incident-postmortem-dag: three independent root tasks fan out at t=0, then a root-cause hypothesizer and postmortem writer synthesize them into one document.
• personalized-interview-simulator: a stateful interviewer (Agent.prompt() across turns) plus a transcript-reading observer, with readline human input and a Zod-validated debrief.

Patterns and integrations

• basics/team-collaboration: runTeam() coordinator pattern.
• patterns/structured-output: any agent returns Zod-validated JSON.
• patterns/multi-perspective-code-review: a generator feeds security, performance, and style reviewers running in parallel, then a synthesizer returns Zod-validated findings.
• patterns/cross-provider-reasoning: preserve a reasoning model's thought stream across a provider switch via preserveReasoningAsText.
• patterns/cost-tiered-pipeline: assign a different model per stage and estimate per-model USD cost from onTrace token counts.
• patterns/fan-out-aggregate: MapReduce-style fan-out via AgentPool.runParallel().
• patterns/agent-handoff: synchronous sub-agent delegation via delegate_to_agent.
• patterns/plan-replay: decompose a goal once with planOnly, serialize it with createPlanArtifact, then replay the same DAG via runFromPlan without re-running the coordinator.
• integrations/trace-observability: onTrace spans for LLM calls, tools, and tasks.
• integrations/mcp-github: expose an MCP server's tools to an agent via connectMCPTools().
• Provider examples: scripts under examples/providers/ covering hosted providers, OpenAI-compatible endpoints, and local models.

Full applications

Clone-and-run apps with their own package.json, not npx tsx scripts. Each embeds OMA in a real backend.

• integrations/express-customer-support: Express REST API. runTasks() behind POST /tickets with per-agent Zod schemas, swappable provider env vars, and HTTP error mapping. Runs on one DeepSeek key (npm install && npm start).
• integrations/with-vercel-ai-sdk: Next.js app. OMA runTeam() plus AI SDK useChat streaming (npm install && npm run dev).

Run any script with npx tsx packages/core/examples/<path>.ts; the full applications above use their own npm scripts.

How is this different from X?

A quick router. Mechanism breakdown follows.

If you need	Pick
Fixed production topology with mature checkpointing	LangGraph JS
Full-stack platform, workflows wired by hand	Mastra
Python stack with mature multi-agent ecosystem	CrewAI
AI app toolkit with broad model-provider support	Vercel AI SDK
TypeScript, goal to result with auto task decomposition	open-multi-agent

vs. LangGraph JS. LangGraph compiles a declarative graph (nodes, edges, conditional routing) into an invokable. open-multi-agent runs a Coordinator that decomposes the goal into a task DAG at runtime, then auto-parallelizes independents. Same end (orchestrated execution), opposite directions: LangGraph is graph-first, OMA is goal-first.

vs. Mastra. Both are TypeScript-native; the difference is who drives the orchestration. With Mastra you wire the workflow by hand. OMA is goal-driven: give its Coordinator a goal and it builds the task DAG at runtime, adapting the plan to the goal instead of running a graph you wired step by step. runTeam(team, goal) in one call.

vs. CrewAI. CrewAI is the mature multi-agent option in Python. OMA targets TypeScript backends with three runtime dependencies and direct Node.js embedding. Roughly comparable orchestration surface; the choice is the language stack.

vs. Vercel AI SDK. AI SDK provides the LLM-call layer — provider abstraction, streaming, tool calls, and structured outputs. It does not orchestrate goal-driven multi-agent teams. The two are complementary: AI SDK for app surfaces and single-agent calls, OMA when you need a team.

Architecture

┌─────────────────────────────────────────────────────────────────┐
│  OpenMultiAgent (Orchestrator)                                  │
│                                                                 │
│  createTeam()  runTeam()  runTasks()  runAgent()  getStatus()   │
└──────────────────────┬──────────────────────────────────────────┘
                       │
            ┌──────────▼──────────┐
            │  Team               │
            │  - AgentConfig[]    │
            │  - MessageBus       │
            │  - TaskQueue        │
            │  - SharedMemory     │
            └──────────┬──────────┘
                       │
         ┌─────────────┴─────────────┐
         │                           │
┌────────▼──────────┐    ┌───────────▼───────────┐
│  AgentPool        │    │  TaskQueue             │
│  - Semaphore      │    │  - dependency graph    │
│  - runParallel()  │    │  - auto unblock        │
└────────┬──────────┘    │  - cascade failure     │
         │               └───────────────────────┘
┌────────▼──────────┐
│  Agent            │
│  - run()          │    ┌────────────────────────┐
│  - prompt()       │───►│  LLMAdapter            │
│  - stream()       │    │  - 13 built-in         │
└────────┬──────────┘    │    providers           │
         │               │  - OpenAI-compatible   │
         │               │  - AI SDK bridge       │
         │               └────────────────────────┘
┌────────▼──────────┐
│  AgentRunner      │    ┌──────────────────────┐
│  - conversation   │───►│  ToolRegistry        │
│    loop           │    │  - defineTool()      │
│  - tool dispatch  │    │  - 6 built-in tools  │
└───────────────────┘    │  + delegate (opt-in) │
                         └──────────────────────┘

Supported Providers

Change provider, model, and set the env var. The agent config shape stays the same.

const agent: AgentConfig = {
  name: 'my-agent',
  provider: 'anthropic',
  model: 'claude-sonnet-4-6',
  systemPrompt: 'You are a helpful assistant.',
}

Kind	How to configure	Services
Built-in shortcuts	Set provider to anthropic, gemini, openai, azure-openai, copilot, grok, deepseek, doubao, hunyuan, minimax, mimo, qiniu, or bedrock; the framework supplies the endpoint.	Anthropic, Gemini, OpenAI, Azure OpenAI, GitHub Copilot, xAI Grok, DeepSeek, Doubao (Volcengine), Hunyuan (Tencent MaaS), MiniMax, MiMo, Qiniu, AWS Bedrock
OpenAI-compatible endpoints	Set provider: 'openai' plus baseURL and, when needed, apiKey.	Ollama, vLLM, LM Studio, llama.cpp server, OpenRouter, Groq, Mistral, Moonshot (Kimi), Qwen, Zhipu
Vercel AI SDK	Import AISdkAdapter from @open-multi-agent/core/ai-sdk; install optional peer ai plus an @ai-sdk/* provider.	Any AI SDK provider (60+ models and hosts)

See docs/providers.md for env vars, model examples, local tool-calling, timeouts, and troubleshooting.

Vercel AI SDK (optional)

Install the optional peer ai plus any @ai-sdk provider you need (for example @ai-sdk/openai). Pass adapter: new AISdkAdapter(model) on AgentConfig to route that agent through the AI SDK instead of the built-in provider factory. provider, apiKey, baseURL, and region are ignored when adapter is set. Mixed teams work as usual: only agents with adapter use the AI SDK.

import { openai } from '@ai-sdk/openai'
import { AISdkAdapter } from '@open-multi-agent/core/ai-sdk'
import { OpenMultiAgent } from '@open-multi-agent/core'

const oma = new OpenMultiAgent()
await oma.runAgent(
  {
    name: 'researcher',
    model: 'gpt-4o',
    adapter: new AISdkAdapter(openai('gpt-4o')),
    systemPrompt: 'You are a researcher.',
  },
  'What are the latest AI trends?',
)

The coordinator accepts the same hook via runTeam(team, goal, { coordinator: { adapter: new AISdkAdapter(...) } }).

Production Checklist

Before going live, wire up the controls that protect token spend, recover from failure, and let you debug.

Concern	Knob	Where it lives
Bound the conversation	maxTurns per agent + contextStrategy (sliding-window / summarize / compact / custom)	AgentConfig
Bound wall-clock time	timeoutMs per agent (aborts a run that hangs, common with local models)	AgentConfig
Cap tool output	maxToolOutputChars (or per-tool maxOutputChars) + compressToolResults: true	AgentConfig and defineTool()
Recover from failure	Per-task maxRetries, retryDelayMs, retryBackoff (exponential multiplier)	Task config used via runTasks()
Survive a crash or restart	checkpoint (pass a runId or a durable MemoryStore) + restore() to resume, skipping completed tasks	OrchestratorConfig / run options
Hard-cap spend	maxTokenBudget on the orchestrator	OrchestratorConfig
Catch stuck agents	loopDetection with onLoopDetected: 'terminate' (or a custom handler)	AgentConfig
Trace and audit	onTrace to your tracing backend; persist renderTeamRunDashboard(result)	OrchestratorConfig
Redact secrets	Automatic — API keys, tokens, and Authorization headers stripped from traces, bash output, and dashboard payloads	built-in (on by default)
Grant tools deliberately	Built-in tools are opt-in (default-deny): an agent gets only what it lists in tools / toolPreset; list neither and it gets none. bash stays unsandboxed once granted, and every tool result is sent to your model provider — so grant read/exec access on purpose. defaultToolPreset restores the old "all tools" behavior in one line	AgentConfig / OrchestratorConfig
Bound filesystem reach	cwd / defaultCwd (default .agent-workspace subdir; widen with process.cwd(), disable with null)	AgentConfig / OrchestratorConfig

Documentation

• Providers — env vars, model examples, local tool-calling, timeouts, troubleshooting.
• Tool configuration — tool presets, custom tools, the filesystem sandbox, and MCP.
• Observability — onProgress events, onTrace spans, and the post-run dashboard.
• Shared memory — the default store and custom MemoryStore backends.
• Checkpoint & resume — opt-in per-run snapshot/resume over any MemoryStore; survive crashes and restarts.
• Context management — sliding window, summarization, compaction, and custom compressors.
• CLI — the JSON-first oma binary for shell and CI.
• Consensus — the runConsensus proposer→judge primitive, the per-task verify hook, and the budget invariant.
• Model routing — the opt-in modelRouting policy: match by phase / agent / role / priority / leaf, first match wins.

Contributing

Issues, feature requests, and PRs are welcome. Some areas where contributions would be especially valuable:

• Production examples. Real-world end-to-end workflows. See examples/production/README.md for the acceptance criteria and submission format.
• Documentation. Guides, tutorials, and API docs.
• Translations. Help translate this README into other languages. Open a PR.

Contributors

Contributor credits by area

Framework features

• @ibrahimkzmv (token budget, context strategy, dependency-scoped context, tool presets, glob, MCP integration, configurable coordinator, CLI, dashboard rendering, trace event types)
• @apollo-mg (context compaction fix, sampling parameters)
• @tizerluo (onPlanReady, onAgentStream)
• @CodingBangboo (planOnly mode)
• @Xin-Mai (output schema validation)
• @JasonOA888 (AbortSignal support)
• @EchoOfZion (coordinator skip for simple goals)
• voidborne-d (OpenAI mixed-content fix, text-tool-extractor depth fix)
• @NamelessNATM (agent delegation base implementation)
• @MyPrototypeWhat (reasoning blocks, reasoning_effort, sampling parity, trace input/output)
• @SiMinus (streaming reasoning events)
• @matthewYang08 (OpenAI reasoning-to-text fallback)
• @dvirarad (OpenAI-family adapter hardening)
• @cat0825 (model routing policy, plan replay, structured shared-memory handoff)
• @mvanhorn (checkpoint & resume)

Provider integrations

• @ibrahimkzmv (Gemini)
• @hkalex (DeepSeek, MiniMax)
• @marceloceccon (Grok)
• @Klarline (Azure OpenAI)
• @Deathwing (GitHub Copilot)
• @JackChiang233 (Qiniu)
• @CodingBangboo (AWS Bedrock)
• @kidoom (MiMo, Doubao)
• @KaitlynFeng (Hunyuan)
• @octo-patch (MiniMax-M3 model upgrade)

Examples & cookbook

• @mvanhorn (research aggregation, code review, meeting summarizer, Groq example, Mistral example)
• @Kinoo0 (code review upgrade)
• @Optimisttt (research aggregation upgrade)
• @Agentscreator (Engram memory integration)
• @fault-segment (contract-review DAG)
• @HuXiangyu123 (cost-tiered example)
• @zouhh22333-beep (translation/backtranslation)
• @pei-pei45 (competitive monitoring)
• @mmjwxbc (interview simulator)
• @binghuaren96 (incident postmortem DAG)
• @DaiMao-UT (paper replication triage)
• @oooooowoooooo (rare disease information triage)
• @CodingBangboo (Express customer support pipeline)
• @nuthalapativarun (Doubao and Zhipu provider examples)
• @goodneamtakenbydogs (Moonshot and Qwen provider examples)
• @suans4746-del (narrative puzzle hint arbitration)
• @gregkonush (Bilig WorkPaper MCP integration)

Docs & tests

• @tmchow (llama.cpp docs)
• @kenrogers (OpenRouter docs)
• @jadegold55 (LLM adapter test coverage)
• @btroops (DeepSeek tool-calling tests)
• @nuthalapativarun (context-management docs)
• @Oxygen56 (errors.ts tests, provider docs for Grok/DeepSeek/Doubao)
• @RheagalFire (LiteLLM gateway docs)

License

MIT