What you can build with OMA.
Runnable recipes straight from the repo — each framed by the problem it solves. Browse by what you want to build, then open the source.
Learn the three ways to run.
Master one agent, coordinator-planned teams, and explicit task DAGs, then compose a mixed-model pool.
One agent with bash and file tools, followed by streaming and multi-turn conversation examples.
run-agenttool-usestreamingUse runTeam() to turn a goal into a coordinated multi-agent task graph and final result.
run-teamtask-dependenciesparallel-executionUse runTasks() with an explicit task DAG, assignments, and dependency-scoped context.
run-taskstask-dependenciesStart from the problem you need to solve.
Complete workflows for support, operations, research, meetings, safety, and other concrete outcomes.
Let runTeam() select the specialists needed for a shipping or billing escalation and synthesize a grounded response.
run-teamRun a four-task contract review DAG with parallel checks and step-level retry.
run-taskstask-dependenciesretryFan three independent incident investigations into root-cause analysis and a final postmortem.
run-taskstask-dependenciesparallel-executionChoose how the work should flow.
Reusable coordination patterns for parallel work, handoffs, consensus, dependencies, and cross-model reasoning.
Send the same question to multiple agents in parallel and aggregate their results.
agent-poolparallel-executionDelegate a synchronous sub-prompt from one specialist to another with delegate_to_agent.
run-teamdelegationGenerate code, run several specialist reviews in parallel, and synthesize structured findings.
run-taskstask-dependenciesparallel-executionUse a proposer and a panel of judges in a configurable refutation and quorum loop.
consensusRun independent research tasks in parallel and collate them through a synthesis task.
run-taskstask-dependenciesparallel-executionPreserve a reasoning model's thought stream as text when a conversation moves between providers.
multi-modelreasoning-transfermulti-turnPut boundaries around the run.
Add schemas, retries, evaluation, budgets, tool gates, plan replay, and observability before a workflow reaches production.
Validate an agent's JSON response against a Zod output schema.
run-agentstructured-outputRetry failed tasks with configurable delay and exponential backoff.
run-tasksretryCombine tool grants with a per-call gate that allows, reviews, or blocks bash commands.
tool-usetool-gatingFreeze a coordinator plan into a diffable artifact and replay it without another planning call.
run-teamtask-dependenciesplan-replayRun a no-key EvalSet regression across two targets with rule and judge scorers plus a gate.
evaluationstructured-outputNo-key examples for batching, trace stores, OpenTelemetry, lifecycle handling, and the offline Run Viewer.
observabilitytrace-storeopen-telemetryEmbed OMA in the systems you already use.
Connect servers, MCP tools, memory, AI SDK applications, external coding agents, and tracing infrastructure.
Embed a structured runTasks() support pipeline behind an Express POST /tickets endpoint.
http-serverrun-tasksstructured-outputConnect an MCP server over stdio and expose its tools through the OMA tool registry.
mcpcustom-toolsCombine OMA runTeam() orchestration with an AI SDK streaming chat interface in Next.js.
ai-sdkrun-teamstreamingRun the same team across your model stack.
A compact directory of provider adapters, OpenAI-compatible endpoints, and local-model setups from the catalog.
Generated from the validated catalog at the pinned Framework commit; all catalog entries are included exactly once. Browse all on GitHub→
Taking this to production?
open-multi-agent is MIT-licensed and free to run yourself. When you need it delivered, integrated, or supported on a deadline, 元定义科技 (YuanASI) offers commercial delivery and support.