AgentTeams coordinates several AI agents in shared chat rooms where people can watch and step in. It uses a Manager-Workers architecture: a Manager agent creates Workers, assigns them tasks, and tracks their progress, so a person directs one agent rather than supervising each Worker individually. AgentTeams does not implement agent logic itself. It orchestrates and manages agent containers, each running an existing agent runtime.
Communication runs over the Matrix protocol. A bundled Tuwunel Matrix server and the Element Web client put the person, the Manager, and the relevant Workers in the same room, so there are no hidden agent-to-agent calls, and any Matrix client (including mobile ones) can join. Workers of different runtime families share a room: OpenClaw, QwenPaw, Hermes, and an experimental DeepSeek Harness worker. A common pattern uses OpenClaw or QwenPaw as a Leader that splits and assigns work, with Hermes Workers executing code.
Credentials stay out of the agents. Workers hold only a consumer token, and the Higress AI gateway keeps the real API keys and GitHub tokens, proxies LLM calls, and hosts MCP servers. MinIO provides a shared file system for exchanging information between agents, which keeps Workers stateless and reduces token use. A Kubernetes-native controller reconciles declarative Worker, Team, Manager, and Human resources.
| OpenClaw native | AgentTeams | |
|---|---|---|
| Deployment | Single process | Distributed containers |
| Agent creation | Manual config and restart | Conversational |
| Credentials | Each agent holds real keys | Workers hold consumer tokens only |
| Human visibility | Optional | Built in (Matrix rooms) |
| Mobile access | Depends on channel setup | Any Matrix client |
| Monitoring | None | Manager heartbeat, visible in the room |
It runs self-hosted on a single machine with Docker or on a Kubernetes cluster through a Helm chart, and works with OpenAI or any OpenAI-compatible model provider, including Qwen.
Features
- Manager-Workers architecture: a Manager agent creates and directs Worker agents, so agents manage other agents
- Multi-runtime rooms: OpenClaw, QwenPaw, Hermes, and experimental DeepSeek Harness Workers collaborate in the same Matrix room
- Human in the loop: every room includes the person, the Manager, and the relevant Workers, and the person can intervene at any point
- Credential isolation: Workers see only a consumer token while the Higress AI gateway holds real API keys and GitHub tokens
- Shared file system: MinIO storage for inter-agent file exchange, reducing token use in multi-agent work
- Built-in IM: a bundled Matrix server and Element Web client, with no DingTalk or Lark bot approval needed
- Declarative resources: a Kubernetes-native controller reconciles Worker, Team, Manager, and Human resources defined in YAML
- Skills on demand: Workers pull community skills from skills.sh and receive custom skills delivered by the Manager
- MCP hosting: MCP servers are declared on Worker, Manager, and Team resources and hosted behind the gateway
- Project tracking: task state history, progress tracking, completion notifications, and an optional dashboard
