# open-multi-agent

> open-multi-agent is an open-source agent framework for multi-agent frameworks. TypeScript multi-agent runtime with durable approvals and verifiable run records. It is available on self-hosted and CLI.

- **Kind:** Agent
- **Pricing:** Open source
- **Platforms:** self-hosted, CLI
- **Website:** https://open-multi-agent.com/go/repo
- **Repository:** https://github.com/open-multi-agent/open-multi-agent
- **GitHub stars:** 7,063
- **License:** MIT
- **Categories:** multi-agent frameworks
- **Listing updated:** 2026-11-14

## About

The library's central idea is that consequential actions pause for a decision. A gate can sit on the plan, a ready task or a tool call; returning suspend persists the request beside the checkpoint, bound to a hash of exactly what the reviewer saw, until another process decides it.

With the run journal on, every block the model saw, tool call and context rewrite is recorded, and verifyRun() rechecks the journal offline. A run can also declare governanceIntent with required roles and be judged on its execution receipt of which roles ran in what order, without reading answer text.

It is self-hosted with no telemetry and no hosted control plane, three runtime dependencies, and default-deny tools. Claude Code, Codex or Gemini CLI can join the task DAG over ACP, and runTeam() lets a coordinator plan a reviewable task graph from a goal.

## Features

- **Durable approvals:** Plan, task or tool-call gates persist until decided.
- **Offline run verification:** verifyRun() checks the journal without a network.
- **Role-based governance:** Judges a run by which required roles executed.
- **Default-deny tools:** Egress policy bounds the built-in adapters.
- **ACP agents in the DAG:** Claude Code, Codex and Gemini CLI can participate.
- **Team and task graphs:** runTeam plans a DAG; runTasks runs one you wrote.

## Good for

- Requiring human sign-off before risky tool calls
- Auditing what a multi-agent run actually did
- Running agent teams in an air-gapped setup
- Mixing coding CLIs inside one workflow

## FAQ

### Does it send telemetry?

No; it states no telemetry and no hosted control plane.

### Which models can it use?

Cloud, local or air-gapped models using your credentials.

### What licence?

MIT.

### Is open-multi-agent free?

Yes. open-multi-agent is open source, so you can use and self-host it for free.

### What are the best alternatives to open-multi-agent?

The closest alternatives to open-multi-agent are hermes-agent, deepseek-harness, langchain, deer-flow and ruflo.

## Alternatives

- [hermes-agent](https://ai-powered.com/agents/hermes-agent): hermes-agent is a freemium open-source agent for ai models and llm tooling.
- [deepseek-harness](https://ai-powered.com/agents/deepseek-harness): deepseek-harness is an open-source agent for multi-agent frameworks.
- [langchain](https://ai-powered.com/agents/langchain): langchain is an open-source agent framework for ai models and llm tooling.
- [deer-flow](https://ai-powered.com/agents/deer-flow): deer-flow is an open-source agent for ai models and llm tooling.
- [ruflo](https://ai-powered.com/agents/ruflo-2): ruflo is an open-source agent framework for knowledge and memory.
- [autogen](https://ai-powered.com/agents/autogen): autogen is an open-source agent framework for multi-agent frameworks.
- [CowAgent](https://ai-powered.com/agents/cowagent): CowAgent is an open-source agent for personal assistants.
- [QwenPaw](https://ai-powered.com/agents/qwenpaw): QwenPaw is an open-source agent for personal assistants.

---
Source: https://ai-powered.com/agents/open-multi-agent
