AgentTeams is an open-source collaborative multi-agent runtime platform. It enables multiple Agents to collaborate in a controlled and auditable room, with full human visibility and intervention capabilities throughout the process.
Built on a Manager-Workers architecture, AgentTeams features a Manager that centrally orchestrates multiple Workers, focusing on collaboration scenarios between humans and Agents, as well as among Agents within enterprise environments.
AgentTeams does not compete with other Agent runtimes. Instead of implementing Agent logic itself, it orchestrates and manages multiple Agent containers (including the Manager and numerous Workers).
<details>
<summary>Using Qwen (通义千问) instead</summary>
</details>
<details>
<summary>Using alternative runtimes (CoPaw Manager + Hermes Workers)</summary>
</details>
<p align="center">
<img src="https://img.alicdn.com/imgextra/i4/O1CN01wHWaJQ29KV3j5vryD_!!6000000008049-0-tps-589-1280.jpg" width="240" />
<img src="https://img.alicdn.com/imgextra/i2/O1CN01q9L67J245mFT0fPXH_!!6000000007340-0-tps-589-1280.jpg" width="240" />
</p>
<p align="center">
<sub>① Manager creates a Worker and assigns tasks</sub>
<sub>② You can also direct Workers directly in the room</sub>
</p>
Key Features
🧬 Manager-Workers Architecture: Eliminates the need for human oversight of individual Worker Claws by enabling Agents to manage other Agents.
🤝 Multi-Runtime Collaboration: OpenClaw, QwenPaw, and Hermes Workers coexist in the same IM room. Use deterministic agents (OpenClaw/QwenPaw) as Leaders to orchestrate tasks, and Hermes Workers for autonomous code execution — each runtime does what it's best at.
📦 MinIO Shared File System: Introduces a shared file system for inter-Agent information exchange, significantly reducing token consumption in multi-Agent collaboration scenarios.
🔐 Higress AI Gateway: Centralizes traffic management and mitigates credential-related risks, alleviating user concerns about security vulnerabilities in the native Lobster framework.
☎️ Element IM Client + Tuwunel IM Server (both Matrix protocol-based): Eliminating DingTalk/Lark integration overhead and enterprise approval workflows. Enables rapid user onboarding to experience the "delight" of model services within an IM environment, while maintaining compatibility with native OpenClaw IM integration.
🧬 IntegrateAgentLoop ：Provides capabilities such as full-stack Agent observability and auditing, Agent evaluation and experimentation, and Agent asset management and continuous optimization.
News
2026-08-08: Release Notes — AgentTeams v1.2.2: adds Manager-to-Worker custom Skill delivery with validation, storage upload, Worker.spec.skills assignment, and QwenPaw hot refresh and enablement without a restart; hardens Manager and Worker Skill loading; and ensures Team Leaders and Workers explicitly join Team Rooms after invitation.
2026-08-06: Release Notes — AgentTeams v1.2.1: unifies the Manager and Worker runtime stack on QwenPaw 2.0.1, publishes the QwenPaw Worker multi-architecture image, strengthens atomic Matrix and TeamHarness task assignment and CoPaw-to-QwenPaw state migration, adds custom model vision and reasoning capability overrides, and improves Worker lifecycle and runtime management reliability.
2026-07-30: Release Notes — AgentTeams v1.2.0 (stable): establishes AgentTeams naming and the final Team/Worker resource contracts end to end; adds the optional AgentTeams Dashboard; and improves Worker storage sync, Team routing and lifecycle convergence, installer support for deploying v1.1.2 with its legacy environment and storage contract (earlier releases still require the matching legacy installer), Dashboard reliability, and tooling and diagnostic safety.
2026-07-17: Release Notes — AgentTeams v1.2.0-beta.1 (prerelease): completes the public rename from the retired predecessor across images, Kubernetes APIs, Helm, Matrix, storage, and runtime contracts; adds the plugin platform, TeamHarness and WorkerFlow integrations, Matrix AppService and Human SSO, model-provider routing and LLM preflight, plus richer controller observability. Beta installation requires explicit opt-in, while the stable default remains v1.1.2.
2026-05-27: Release Notes — AgentTeams v1.1.2: QwenPaw-first installer with keep-all upgrade flow, Team human coordinators and refreshed Team Leader coordination tools, Nacos remote skills with sts-agentteams / ai-registry STS scope, Worker CR-name decoupled from runtime name, controller reconcile metrics and graceful shutdown.
2026-05-07: Release Notes | Changelog — AgentTeams v1.1.1: declarative MCP on Worker/Manager/Team CRDs (breaking) and on Team Leader, custom spec.env for CRs, Token Plan + Qwen Cloud international + qwen3.6-plus, namespace-scoped controller RBAC, optional SOUL.md in Worker packages.
2026-04-14: English | 中文 — Deep dive: AgentTeams as a Kubernetes-native multi-agent collaboration orchestration system.
2026-04-03: English | 中文 — AgentTeams 1.0.9: Kubernetes-style declarative resource management (YAML for Worker, Team, Human); Worker Template Marketplace; Manager QwenPaw runtime; Nacos Skills Registry and more.
2026-03-14: English | 中文 — AgentTeams 1.0.6: enterprise-grade MCP Server management, zero credential exposure.
2026-03-10: English | 中文 — AgentTeams 1.0.4: QwenPaw (formerly CoPaw) Worker support, 80% less memory.
2026-03-04: English | 中文 — AgentTeams open sourced under its former name.
Why AgentTeams
Enterprise-Grade Security: Worker Agents operate with consumer tokens only. Real credentials (API keys, GitHub PATs) stay in the gateway — Workers can't see them, and neither can attackers.
Fully Private: Matrix is a decentralized, open protocol. Host it yourself, federate with others if you want. No vendor lock-in, no data harvesting.
Human-in-the-Loop by Default: Every Matrix room includes you, the Manager, and the relevant Workers. Watch everything. Jump in anytime. No black boxes.
Zero Configuration IM: Built-in Matrix server means no bot applications, no API approvals, no waiting. Just open Element Web and start chatting.
One Command Setup: curl | bash and you're done — AI gateway, Matrix server, file storage, web client, and Manager Agent.
Skills Ecosystem: Workers pull from skills.sh (80,000+ community skills) on demand. Safe because Workers can't access real credentials.
Quick Start
Prerequisites: Docker Desktop (Windows/macOS) or Docker Engine (Linux).
Resources: 2 CPU cores + 4 GB RAM minimum. For multiple Workers, 4 cores + 8 GB recommended.
Set-ExecutionPolicy Bypass -Scope Process -Force; $wc=New-Object Net.WebClient; $wc.Encoding=[Text.Encoding]::UTF8; iex $wc.DownloadString('https://raw.githubusercontent.com/agentscope-ai/AgentTeams/main/install/agentteams-install.ps1')
The installer walks you through:
Choose your LLM provider (OpenAI-compatible APIs supported)
Enter your API key
Select network mode (local-only or external access)
Wait for setup to complete
Access
Open http://127.0.0.1:18088 in your browser and log in to Element Web. The Manager will greet you and explain how to create your first Worker.
Mobile: Use any Matrix client (Element, FluffyChat) and connect to your server address.
That's it. No bot applications. No external services. Your AI team runs entirely on your machine.
Upgrade
# Upgrade to latest (preserves all data)
bash <(curl -sSL https://raw.githubusercontent.com/agentscope-ai/AgentTeams/main/install/agentteams-install.sh)
# Upgrade to specific version
AGENTTEAMS_VERSION=v1.2.2 bash <(curl -sSL https://raw.githubusercontent.com/agentscope-ai/AgentTeams/main/install/agentteams-install.sh)
This removes all AgentTeams containers (Manager, Workers, docker-proxy), Docker volume, network, env file, workspace directory, and install log.
Install on Kubernetes (Helm)
For shared / production deployments you can install AgentTeams on any Kubernetes cluster via the official Helm chart. The default profile bundles the Higress AI gateway, Tuwunel (Matrix), MinIO and the AgentTeams controller — no external dependencies required.
The commands below provide a quick installation path. See the Kubernetes Deployment Guide for cluster planning, values files, model services, runtimes, persistence, Ingress, and operations.
| Value | Required | Description |
|---|---|---|
| credentials.llmApiKey | yes | API key for your LLM provider |
| gateway.publicURL | yes | Public URL where users will reach Element Web (e.g. http://localhost:18080 for port-forward, or https://agentteams.example.com for an Ingress) |
| credentials.adminPassword | recommended | Matrix admin password; auto-generated if left empty (you'll have to read it back from the Secret) |
| credentials.llmProvider | no | LLM provider name, defaults to openai-compat |
| credentials.defaultModel | no | Default model, defaults to gpt-5.4 |
| credentials.llmBaseUrl | no | OpenAI-compatible base URL (e.g. https://api.deepseek.com/v1). Leave empty for official OpenAI API |
| preflight.llm.enabled | no | Run an install/upgrade hook that validates the LLM API key, base URL, and model before the controller starts. Defaults to true |
| preflight.llm.strict | no | Fail the Helm install/upgrade when the LLM preflight fails. Defaults to true; set to false to emit a warning and continue |
| preflight.llm.timeoutSeconds | no | Per-request timeout for the LLM preflight HTTP probe. Defaults to 30 |
| preflight.llm.retries | no | Retry count for transient LLM preflight failures such as rate limits and provider 5xx responses. Defaults to 2 |
| preflight.llm.activeDeadlineSeconds | no | Kubernetes Job active deadline for the preflight hook. Defaults to 120 |
| preflight.llm.resources | no | Optional Kubernetes resource requests/limits for the preflight hook container |
| manager.runtime | no | Manager runtime: OpenClaw uses openclaw (default); CoPaw uses qwenpaw in the current chart, with copaw retained as a compatibility alias. Managers do not support Hermes |
| worker.defaultRuntime | no | The chart provides default image values for openclaw (default), copaw, hermes, and openhuman, but the current CRD rejects an explicit spec.runtime: openhuman; set spec.image explicitly for a QwenPaw Worker |
Helm installs run an LLM preflight hook by default. The hook sends a minimal OpenAI-compatible /chat/completions request using credentials.llmApiKey, credentials.llmBaseUrl, and credentials.defaultModel; invalid keys, unreachable base URLs, unsupported models, quota errors, and provider outages fail the install before the controller starts. To bypass this check for restricted or offline clusters:
When selecting the CoPaw Manager, set both manager.runtime=qwenpaw and the agentteams-manager-qwenpaw image. Configure Worker default images through worker.defaultImage.<runtime>.
Multi-Region Image Registry
Images point to the China region by default. When deploying outside China, use a nearby registry for faster pulls:
| Region | Registry |
|---|---|
| China (default) | higress-registry.cn-hangzhou.cr.aliyuncs.com/higress |
| North America | higress-registry.us-west-1.cr.aliyuncs.com/higress |
| Southeast Asia | higress-registry.ap-southeast-7.cr.aliyuncs.com/higress |
global.imageRegistry affects only subcharts that consume that global value. The Controller, Manager, Workers, Tuwunel, MinIO, and Element Web use their own complete image.repository values. Override every relevant image in a values file when switching regions or using a private registry. See the Kubernetes Deployment Guide and helm/agentteams/values.yaml for the complete configuration.
Access
For a temporary local admin session, forward the Higress Gateway:
Then open http://localhost:18080 and log in to Element Web. The port-forward
ends when the command exits and is not suitable for shared access.
For company or Internet access, expose only svc/higress-gateway through an
HTTPS Ingress or LoadBalancer. gateway.publicURL is written into the Element
Web configuration as its Matrix homeserver URL, so it must exactly match the
public origin that users open (for example, https://agentteams.example.com).
Point the public DNS name at your Ingress controller or load balancer.
Provision a TLS certificate Secret in agentteams-system.
Route that host to the Higress Gateway. This generic example assumes an
NGINX IngressClass and an existing agentteams-tls Secret; replace both to
match your cluster:
Keep the controller API, Tuwunel, MinIO, and the Higress Console private unless
you apply separate authentication and network policy. Use HTTPS for shared
access because Matrix login credentials and access tokens pass through this
endpoint. A LoadBalancer Service can replace the Ingress, but the DNS, TLS,
and gateway.publicURL requirements remain the same.
You: Create a Worker named alice for frontend development
Manager: Done. Worker alice is ready.
Room: Worker: Alice
Tell alice what to build.
You: @alice implement a login page with React
Alice: On it... [a few minutes later]
Done. PR submitted: https://github.com/xxx/pull/1
Security Model
Worker (consumer token only)
→ Higress AI Gateway (holds real API keys, GitHub PAT)
→ LLM API / GitHub API / MCP Servers
Workers see only their consumer token. The gateway handles all real credentials. The Manager knows what Workers are doing but never touches the actual keys.
Human in the Loop
Every Matrix Room includes you, the Manager, and relevant Workers:
You: @bob wait, change the password rule to minimum 8 chars
Bob: Got it, updated.
Alice: Frontend validation updated too.
No hidden agent-to-agent calls. Everything is visible and intervenable.
Multi-Runtime Collaboration
AgentTeams currently provides three main Worker runtime families that can coexist in the same IM room, collaborating on tasks together:
OpenClaw (Node.js) — General-purpose agent with rich skills ecosystem, ideal for task orchestration and tool calling
QwenPaw (Python) — Lightweight runtime, suited for browser automation and quick tasks
Hermes (hermes-agent) — Autonomous coding agent with terminal sandbox, self-improving skills, and persistent memory
The copaw value remains available for compatibility with existing CoPaw Workers; the current CRD accepts both copaw and qwenpaw. Each runtime excels at different tasks. A common pattern is to use deterministic agents (OpenClaw/QwenPaw) as Leaders to decompose and assign work, and Hermes Workers for autonomous code execution. All runtimes communicate through Matrix rooms—fully visible and fully intervenable.
# Switch any worker's runtime in place
agt update worker --runtime hermes
Architecture
┌───────────────────────────────────────────────┐
│ agentteams-controller │
│ Higress │ Tuwunel │ MinIO │ Element Web │
└──────────────────┬────────────────────────────┘
│ Matrix + HTTP Files
┌──────────────────┴──────────┐
│ agentteams-manager-agent │
│ Manager (OpenClaw/ │
│ QwenPaw) │
└──────────────────┬──────────┘
│
┌──────────────────┼────────────────────────────┐
│ │ │
▼ ▼ ▼
Worker Alice Worker Bob Worker Charlie
(OpenClaw) (QwenPaw) (Hermes)
| Component | Role |
|-----------|------|
| agentteams-controller | Kubernetes-native control plane, reconciles Worker/Team/Manager CRs |
| Higress AI Gateway | LLM proxy, MCP Server hosting, credential management |
| Tuwunel (Matrix) | Self-hosted IM server for all Agent + Human communication |
| Element Web | Browser client, zero setup |
| MinIO | Centralized file storage, Workers are stateless |
AgentTeams vs OpenClaw Native
| | OpenClaw Native | AgentTeams |
|---|---|---|
| Deployment | Single process | Distributed containers |
| Agent creation | Manual config + restart | Conversational |
| Credentials | Each agent holds real keys | Workers only hold consumer tokens |
| Human visibility | Optional | Built-in (Matrix Rooms) |
| Mobile access | Depends on channel setup | Any Matrix client, zero config |
| Monitoring | None | Manager heartbeat, visible in Room |
Then open the AgentTeams repo in Cursor, Claude Code, or similar AI tool and ask:
"Read the JSONL files in debug-log/. Analyze the Matrix message logs and agent session logs together. Cross-reference with the AgentTeams codebase to identify the root cause of [describe your bug]."
You can also let the AI tool submit the issue or PR directly. Install GitHub CLI, run gh auth login to authenticate in your browser, then add the OpenClaw GitHub skill to your AI coding tool (Cursor, Claude Code, etc.). After that, just ask it to file the issue or open a PR based on its analysis.
Build & Test
make build # Build all images
make test # Build + run all integration tests
make test-quick # Smoke test only
Other Commands
make replay TASK="Create a Worker named alice for frontend development"
make uninstall
make help